blitzstrike 2.4.139 → 2.4.140

package/CHANGELOG.md

@@ -5,6 +5,19 @@ All notable changes to Blitz Strike are documented here.
 The format follows [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/).
 
+## [2.4.140] — 2026-10-02
+
+### Added — severity ≤ evidence ceiling (Decepticon SPEC V4)
+
+- `severity_calibrate` now takes an `evidence_level` (declared | resolved |
+  call_path | runtime) that CAPS the reportable severity + priority: declared ≤
+  low, resolved ≤ medium, call_path ≤ high, runtime ≤ critical. The raw derived
+  severity/CVSS is retained; only the REPORTED severity is capped.
+- New `evidenceCeiling(level)` deterministic ladder + doctrine rule
+  `severity_evidence_ceiling`: never claim "critical" on a code-only hit that was
+  never executed. Wired into severity_calibrate + system prompt. 23 doctrine rules,
+  515/515 tests.
+
 ## [2.4.139] — 2026-10-02
 
 ### Added — coverage-completeness doctrine (dead verdict = confirmed)

package/dist/index.js

@@ -60068,11 +60068,35 @@ var VRT_MAP = {
   toctou: { vrt: "insecure_design > race_condition", priority: "P2", severity: "high", rationale: "A check-then-act gap lets an attacker race the operation — balance double-spend, symlink swap, auth bypass (CWE-367)." },
   second_order: { vrt: "injection > second_order_injection", priority: "P2", severity: "high", rationale: "Stored data rendered/queried without escaping — stored XSS or second-order SQLi (CWE-79/89)." }
 };
+var SEVERITY_RANK = {
+  informational: 0,
+  low: 1,
+  medium: 2,
+  high: 3,
+  critical: 4
+};
+var SEVERITY_PRIORITY = {
+  informational: "P5",
+  low: "P4",
+  medium: "P3",
+  high: "P2",
+  critical: "P1"
+};
+var EVIDENCE_CEILING = {
+  declared: "low",
+  resolved: "medium",
+  call_path: "high",
+  runtime: "critical"
+};
+function capSeverity(severity, ceiling) {
+  return SEVERITY_RANK[severity] <= SEVERITY_RANK[ceiling] ? severity : ceiling;
+}
 function severityCalibrate(input) {
   const type = input.type?.toLowerCase() ?? "";
   const entry = VRT_MAP[type];
+  let result;
   if (input.informational) {
-    return {
+    result = {
       type,
       vrt_category: "informational",
       priority: "P5",
@@ -60080,9 +60104,8 @@ function severityCalibrate(input) {
       rationale: "Marked informational — excluded from the validity ratio.",
       ...input.cvss != null ? { cvss_severity: cvssSeverity(input.cvss) } : {}
     };
-  }
-  if (!entry) {
-    return {
+  } else if (!entry) {
+    result = {
       type,
       vrt_category: "unclassified",
       priority: "P4",
@@ -60090,15 +60113,29 @@ function severityCalibrate(input) {
       rationale: `No VRT mapping for '${type}'; defaulting to low until classified.`,
       ...input.cvss != null ? { cvss_severity: cvssSeverity(input.cvss) } : {}
     };
+  } else {
+    result = {
+      type,
+      vrt_category: entry.vrt,
+      priority: entry.priority,
+      severity: entry.severity,
+      rationale: entry.rationale,
+      ...input.cvss != null ? { cvss_severity: cvssSeverity(input.cvss) } : {}
+    };
   }
-  return {
-    type,
-    vrt_category: entry.vrt,
-    priority: entry.priority,
-    severity: entry.severity,
-    rationale: entry.rationale,
-    ...input.cvss != null ? { cvss_severity: cvssSeverity(input.cvss) } : {}
-  };
+  if (input.evidence_level) {
+    const ceiling = EVIDENCE_CEILING[input.evidence_level] ?? "low";
+    if (SEVERITY_RANK[result.severity] > SEVERITY_RANK[ceiling]) {
+      const capped = capSeverity(result.severity, ceiling);
+      result = {
+        ...result,
+        severity: capped,
+        priority: SEVERITY_PRIORITY[capped],
+        rationale: `${result.rationale} [capped to ${capped}: evidence level '${input.evidence_level}' supports at most ${ceiling}]`
+      };
+    }
+  }
+  return result;
 }
 function listSeverityMap() {
   return Object.entries(VRT_MAP).map(([type, e]) => ({ type, priority: e.priority, vrt: e.vrt }));
@@ -62865,6 +62902,11 @@ var HUNTING_DOCTRINE = [
     id: "coverage_completeness",
     title: "Coverage Completeness — A Dead Verdict Is Worth a Confirmed One",
     rule: "Every class × eligible shape must carry a terminal verdict: confirmed, dead (controlled disproof), partial (blocked prerequisite), or skipped (explicit policy exclusion). A dead verdict backed by controlled evidence is worth as much as a confirmed one — record it in the ledger so goal cycles never re-test what you already disproved. COMPLETE means zero shapes without a terminal verdict, not 'found N bugs'; breadth of verdicts is the deliverable, not a bug count."
+  },
+  {
+    id: "severity_evidence_ceiling",
+    title: "Severity ≤ Evidence Ceiling",
+    rule: "The severity a finding may CLAIM is capped by its demonstrated-impact evidence: declared (found in code/name) ≤ low, resolved (symbol resolved but not called) ≤ medium, call-path (reached via call graph) ≤ high, runtime (actually exploited) ≤ critical. Report the capped severity via severity_calibrate(type, cvss, evidence_level); keep the raw CVSS/EPSS/KEV as reference. Never claim 'critical' on a code-only hit that was never executed."
   }
 ];
 function huntingDoctrine() {
@@ -63648,7 +63690,7 @@ function createServer2() {
 ` + `- A hard block (Cloudflare/WAF/IP-block/captcha) is NOT the end: mark the finding blocked + note the required vantage (residential / authorized / correct-SNI / --resolve direct-to-origin), then CONTINUE to the next lead (origin bypass, non-prod cluster, mobile APK, subdomain).
 ` + `- Keep trying every lead AND every chain step until each is verified/executed or blocked, THEN generate_report. Never deliver a bare surface map, and never stop at 'found X' when X enables more. Write the report autonomously — when findings are verified (or blocked), call generate_report immediately; never ask 'should I write the report now?'. The report is the automatic end-state of the engagement, not a permission-gated step.
 
-` + "TWO PATHS:\\n" + "- Source path on disk -> blitz_scan / taint_file / eagle_eye2 (static).\\n" + "- Live URL -> live_recon / active_scan / strike_verify (web-hunting).\\n\\n" + "DEEPEN A FINDING: attack_plan(target, tech, params) to get the prioritized vector plan (decide what to test next with data); list_chains (escalation), chain_links (compose), technique_lookup(cls), payload_lookup, bypass_lookup (blocked). Ground any detector hit empirically with pattern_lookup(id) — it returns the class's CWE, real-world signature, escalation chains, citable references, per-class bypass techniques, and the high-value assets (crown jewels) that class hits, so a raw hit becomes a class-aware finding. Resolve free-form terminology first with synonym_lookup(term) (IDOR/BOLA/SQLi/prompt-injection → canonical id). For an llm_injection hit, its pattern carries a validation discipline — the RUN-TWICE RULE (reproduce token-for-token across two fresh sessions), NON-GUESSABLE ANCHOR (must leak a real key/URL, not a plausible guess), and OOB PROOF (a verifiable callback log) — apply all three before reporting. BEFORE generating a report, run submission_gate(...) and only ship a finding that passes all 10 questions (reproducible, in-scope, real-impact, no-privileged-assumption, verified-live, evidence-redacted, severity-derived, impact-demonstrated, production-delivery, no-custom-headers); one false = kill it. TRIAGE HONESTY (apply before you ship): credential liveness ≠ impact — probe AUTHORITY (capability ceiling / denied-op control / metadata-introspection), not just 'the token works'; harness proof ≠ attacker delivery — prove the external path through normal flow, not a victim-value harness; severity = max DEMONSTRATED, not hypothetical ceiling; classify intent first (DOCUMENTED AND MATCHES = reject, DOCUMENTED BUT EXCEEDED = report the excess, CONTRADICTED = strong evidence, NOT DOCUMENTED = don't call it intended) — an unauth response is not automatically a bug. When a PoC/report comes out under-enriched, wet_dry(type, wet, dry) diagnoses WHERE it broke — WET (model/prompt/timeout) vs DRY (parser) vs template. Generate the concrete reproduction recipe with auto_poc(type, base_url, endpoint, param) — it emits the exact marker request + negative control + expected differential + Patchstack-style title, so 'reproducible' is executable, not abstract. Calibrate severity with severity_calibrate(type, cvss) to get a P1–P5 tier + VRT-style category. Start each engagement with engagement_scaffold(target, scope) for the folder + scope.md structure, and check data_refresh() to confirm the data layers are consistent. Phase doctrine: read_skill(bs-orchestrate-engagement | bs-web-hunting | bs-source-audit | bs-verify-finding). Campaign plan: read_playbook(target-type). On any failure call retry_guidance(signal) / model_fallback(signal) before giving up. HUNTING DOCTRINE (maximize payout rate): (1) SIBLING RULE — if one endpoint requires auth, check every sibling (/export /delete /share); ~30% of paid IDOR/auth bugs come from siblings — AUTOMATE it with sibling_scan(base_url, endpoint, token_a, token_b), the deterministic live logic-bug prober (sibling enumeration + differential IDOR + auth matrix + mass assignment + rate limit, evidence-first). (2) A→B SIGNAL — a confirmed bug signals a class of developer mistake; hunt B and C (20-min box) before reporting. (3) TWO-ACCOUNT IDOR — always test with attacker account A reading victim account B, and state it in the report. (4) PoC ESCALATION — IDOR shows victim data (not 200), XSS shows cookie exfil (not alert), SSRF shows internal response (not DNS), SQLi shows DB content (not error). (5) FOLLOW THE MONEY — billing/credits/refunds/wallet flows. (6) Kill weak findings with kill_list(type) BEFORE the gate — missing headers, self-XSS, open-redirect-alone, SSRF-DNS-only, GraphQL-introspection-alone, pure-DoS, SSRF-below-High, session-swap, low-sensitivity-read-only, self-only-state are always N/A standalone. (7) CI/CD + SAML/SSO are critical attack surfaces (pull_request_target, expression injection, XML signature wrapping, comment injection). (8) ACCESS MODE ROUTING — derive the lane from ACTUAL auth state (2 accounts = bidirectional IDOR/A→B, 1 = single-account, 0 = UNAUTH); UNAUTH is not failure — bundle/APK secrets, exposed APIs, admin/metrics/GraphQL, and supply-chain checks all run unauth. (9) NO CUSTOM HEADERS — a real attacker sends no X-Bug-Bounty / X-Forwarded-For 'just in case' / program-mandated throttling; findings that only fire WITH researcher headers are known issues, not attack surface (auth + explicit test-payload headers are the payload itself and allowed). (10) FUNCTIONAL SIBLING SCOPE — a first-party sibling directly used by the app (auth.X / account.X / api.X / login.X for signup/login/account/API/billing) is IN SCOPE even when display names only www.X; third-party IdP/captcha/CDN/payment = external; require flow evidence (redirect/link/form-action/API), and an explicit program exclusion always wins. (11) COVERAGE COMPLETENESS — every class × eligible shape must carry a terminal verdict (confirmed / dead / partial / skipped); a DEAD verdict with controlled evidence is worth as much as a confirmed one (record it so it is never re-tested), and COMPLETE = zero shapes without a terminal verdict, not 'found N bugs'. Full doctrine: hunting_doctrine(). WHAT TO HUNT NEXT: watchlist() returns the ranked CWE frequency + high-value CVEs with payout + a saturated flag (prioritize high-freq/high-payout/not-saturated: missing-authz, SQLi, SSRF, deserialization; skip saturated XSS/CSRF/open-redirect unless they chain into takeover). The CVE watchlist now spans 4,000+ product fingerprints (live from nuclei-templates when installed — refresh with sync_cve_corpus() before a big recon pass); live_recon's suggested_cves names the SPECIFIC CVE for the detected product/version — then cve_verify(cve, target) CONFIRMS it live by running the matching nuclei template (VERIFIED only when nuclei actually matches; a CVE is report-ready only after cve_verify, never from the fingerprint alone). For 'scan this target' automation, nuclei_scan(target) fingerprints the stack and runs the WHOLE matched template set (CVE templates + optional broad tech-tag) in one rate-limited pass — use it instead of verifying CVEs one-by-one. SURFACE-FIRST: surface_map(surface) maps an attack surface (image/redirect proxy, edge cache, payment portal, AI agent, file service, Electron, SSO) to its highest-value tests — map surfaces then test, don't spray payloads; then param_surface(target, class) discovers EVERY testable parameter (URL query, form fields, /users/123 path ids + UUIDs, JS refs) + per-class specialist hints — test the DISCOVERED params, not just the hinted one. API RECON (you reason this — no tool needed): when discoverApi/live_recon surfaces an OpenAPI/Swagger spec (/openapi.json /swagger.json /api-docs), FETCH and PARSE it yourself (plain JSON/YAML) — extract every endpoint + method + parameter + auth requirement, map the object-level paths ({id}/{uuid}) to BOLA/IDOR candidates, then test them with sibling_scan. OAuth/SSO flows are also your reasoning job — follow the redirect/state sequence manually. PRECISION GUARD: resolve_check(host) detects when an ISP/country redirects the domain to a blocking/sinkhole page (gambling/adult/streaming niches) — audit the PUBLIC IP, never the sinkhole. CIRCUIT BREAKER: target_health(target) probes the target and tracks consecutive failures — if the circuit OPENS (5 consecutive timeouts/5xx) PAUSE the target and resume when it recovers (auto-closes on a successful probe); circuit_status() lists every tracked target's state.\n" + "KNOWLEDGE BASE (discover, don't guess): skillLookup(topic) / listSkills for the 40 attack-path + hunting skills (ap-*, hp-*, bs-*); list_attack_vectors / attack_vectors(cat) for the 34-category vector taxonomy; taxonomy(kind) for OWASP/CWE/ASVS/API-top-10; read_playbook(type) for the 17 campaign playbooks; list_manuals / read_tool_manual for 317 tool references; list_tools / run_catalog_tool for the 141-tool catalog. Every tool, skill, and data layer is reachable from these. CROSS-TARGET INTELLIGENCE: strike_resolve auto-records confirmed/false_positive verdicts into the empirical ledger; query it with intelligence(action=hit_rates/top_vectors/recall) so a new target on a known tech inherits historical hit-rates (attack_plan already blends them into success_probability). intelligence(action=recall, tech, target) is the per-engagement pattern recall — it returns which patterns historically CONFIRMED on that tech (prioritize them) versus which only ever produced false positives (avoid them), so a new engagement starts from the empirical playbook of past ones. LIVE PENTEST + TARGETING (black-box, WAF-safe): cache_gap(base_url, endpoint) fuzzes path-normalization + static-extension variants for cache deception/poisoning; blind_oracle(base_url, param, value) detects which SINK a param reaches (sql/ssti/xss) via benign arithmetic (id=1+0, test${7*7}) WITHOUT a malicious payload; param_precedence(base_url, param, A, B) maps duplicate-parameter precedence (first/last-wins) for WAF-bypass; broken_link_scan(url) detects expired-domain takeover (crawls outbound refs → DNS/RDAP → flags claimable domains); asset_discovery(domain) enumerates the FULL passive attack surface — crt.sh subdomains + DNS brute + A-record resolution + optional port scan. Logic-bug live probers: jwt_forge(token, endpoint, …) forges JWT confusion tokens (alg:none / RS256→HS256 key-confusion / kid-jku-x5u) and proves which is ACCEPTED live; proto_pollution_probe(target, endpoint) sends __proto__ / constructor.prototype JSON canaries and reports VERIFIED only on reflection; smuggling_test(target) sends CL.TE / TE.CL desync probes over a raw connection; totp(secret) generates an RFC 6238 TOTP token so the agent can fill a 2FA field itself; auth_login(login_url, user, pass, …) POSTs credentials and extracts the session cookie/token for authenticated scanning (use browser_agent navigate/type/click/get_cookie for JS-heavy logins); screenshot(url) captures a page image for vision-based proof. OOB (out-of-band) PROOF for BLIND vulns: oob_start() spins an interactsh listener (unique *.oast.* domain, streams DNS/HTTP interactions — the standard for public targets) or a local HTTP listener for local labs; inject the canary, then oob_poll(id) until the hit arrives; oob_stop(id) to clean up. A blind SSRF/XXE/SQLi/command-injection is CONFIRMED only when the canary actually arrives — never from a guess. SOURCE TARGETING: review_harness(path) FIRST for full coverage — it applies a fixed 5-signal rubric (entry_points/auth_gates/sinks/secrets/logic_bugs) to EVERY source file and sorts the signal-dense ones first (a complete map, not a partial scan); then action_surface(code) maps every wp_ajax_* action to its vuln class + DERIVED priority (what to look for WHERE); security_state(sanitizers, sink) gives a wrong-context sanitization verdict; framework_knowledge(framework) returns auth/nonce primitives for 9 frameworks; variant_scan(root, signature) mass-sweeps an install base for ONE bug family; patch_analyze(old, new) REVERSES a security patch into the 0-day, and its mine_signature feeds variant_scan to sweep unpatched copies. All of these are DETECTION HINTS — verify each live (strike_verify / browser_validate) before it enters the report. NEWEST DETECTORS — understand AND try them: variant_scan also emits priv_esc (privilege escalation — a role/capability/user change fed by request input, traced cross-file via reverse call-graph, e.g. set_role($id, sanitize_key($_POST['um-role'])) or update_user_meta(...,'wp_capabilities', $_POST[...])) and flags uploadAllow=>array('all') unrestricted-upload configs as file_upload (the file-manager/connector shape). TO TRY a priv_esc hit: submit the role field with 'administrator' (or send wp_capabilities meta) and confirm the account actually gains admin. TO TRY an unrestricted-upload hit: POST a .php/.phtml and confirm it lands web-reachable + executes. variant_scan also emits hardcoded_secret (CWE-798) in three tiers: a secret field (api_key/db_password/access_token/aws_*) assigned a literal (HIGH), a known secret FORMAT (JWT eyJ…/AWS AKIA…/GitHub ghp_…/Slack xoxb-…/…) regardless of name (HIGH), or a high-entropy suspect token (MEDIUM — verify before reporting). Verify a frontend/JS secret live via drive_devtools (inspect the JS bundle / network response and confirm the key is actually exposed), and for a backend secret confirm it is reachable/valid. A static hit is a HYPOTHESIS until verified live — never report unverified. For a repo, also scan_git_history(root) — secrets DELETED from HEAD still live in .git history (the removed-later-is-not-fixed trap), and it feeds every recovered pre-deletion blob through the same secret detector. Once you recover a credential, prove it is still ACTIVE with validate_leaked_key(apiBase, endpoints, key|blobUrl) — a READ-ONLY differential test (401-without-key vs 200-with-key = auth bypass PROVEN). Two supply-chain detectors: dependency_confusion (a manifest/build file resolving a package from a public registry — pypi.org/npmjs.org --extra-index-url, or a scoped @company/pkg with no private registry) and ml_supply_chain (torch.load without weights_only=True / joblib.load / np.load(allow_pickle=True) / keras load_model on an untrusted model → pickle RCE). Rust is supported end-to-end: the taint engine (blitz_scan / taint_file) maps Rust sinks (std::process::Command, sqlx/diesel, reqwest/hyper, std::fs, serde/bincode) to the classic classes, plus two Rust-specific detectors — rust_unsafe (transmute / assume_init / from_raw_parts / ptr::* — memory-unsafety surface, CWE-119) and rust_format_string (println!/format! with a variable as the format string — CWE-134). The sqlx query! compile-time macro is treated as safe (suppressed). Two more web-class detectors: jwt_alg_confusion (a JWT verifier that dispatches on the token's alg header and accepts BOTH RS256 and HS256, with the HS256 HMAC secret derived from public/shared key material — CWE-347; the attacker sets alg=HS256 and signs with the public key to forge admin tokens) and cache_deception (a reverse proxy caching static-extension URLs with a cache key that omits the session cookie, over a session-authenticated backend — CWE-525; the victim's cached private response is re-fetched anonymously). TO TRY a jwt_alg_confusion hit: fetch the public key (/.well-known/jwks.json), forge an HS256 token signed with that public key, and confirm admin access. TO TRY a cache_deception hit: request an authenticated path with a static extension appended (.css/.js) and re-fetch it anonymously. Three more detectors: cors_misconfiguration (Access-Control-Allow-Origin reflects the request Origin or is '*' while credentials are true — CWE-942; any site reads the victim's authenticated responses), xss (user input into a DOM sink — innerHTML / dangerouslySetInnerHTML / document.write / .html() / eval — CWE-79), and subdomain_takeover (a dangling CNAME/AliasTarget to a claimable service like *.github.io/*.herokuapp.com/*.amazonaws.com — CWE-404). TO TRY a cors hit: send an evil Origin + confirm the response echoes it with Allow-Credentials:true. TO TRY an xss hit: confirm the sink executes (drive_devtools). TO TRY a takeover hit: verify the CNAME target resolves to an unclaimed/not-found page. Two more client-side detectors: postmessage_no_origin_check (a window message listener that reads event.data but never validates event.origin — CWE-345; any site can drive the handler) and client_trusted_flag (a privilege/entitlement flag like isAdmin/withVip/premium read from request input and used in a security decision — CWE-807/639; derive the flag from server state, never the client). TO TRY a postMessage hit: send a crafted message from an attacker origin and confirm the handler acts on it. TO TRY a client-trusted-flag hit: flip the flag param to true and confirm the privilege is actually granted. Four newest detectors: open_redirect (user input into a redirect/Location sink — CWE-601; TO TRY: supply a marker redirect URL and confirm the 30x Location reflects it), missing_nonce/CSRF (a state-changing wp_ajax/form/Express handler with no nonce or CSRF token — CWE-352; TO TRY: replay the state-changing request without the token and watch the state change), toctou (a check-then-act gap like file_exists→unlink or balance-check→withdraw — CWE-367; TO TRY: race the gap with concurrent requests via race_test), and second_order (a DB read reaching echo/query/eval without escaping — stored XSS or second-order SQLi that single-pass source→sink misses; TO TRY: seed the stored field with a marker and observe it render). WEB3: for Solidity/EVM, call web3_audit(code|path) — a deterministic 13-class smart-contract audit (reentrancy incl. ERC721/ERC1155 callback reentrancy, unchecked return, unchecked arithmetic, tx.origin auth, signature replay, unprotected selfdestruct, delegatecall-to-user-input, missing access control, timestamp dependence, spot-price oracle manipulation incl. slot0/latestRoundData, unbounded loop, abi.encodePacked hash collision, missing address(0) validation) returning ranked findings (class/CWE/severity/line), then foundry_poc(class, contract) to generate an executable Foundry test, and foundry_run(contract_code, poc_code|class_) to actually EXECUTE it — a PASSING PoC test is the proof that the exploit reproduces (deterministic: forge's own test verdict). A web3 hit is a HYPOTHESIS until the Foundry PoC reproduces it.\n" + `PRECISION ANALYSIS — chrome-devtools MCP is MANDATORY, not optional:
+` + "TWO PATHS:\\n" + "- Source path on disk -> blitz_scan / taint_file / eagle_eye2 (static).\\n" + "- Live URL -> live_recon / active_scan / strike_verify (web-hunting).\\n\\n" + "DEEPEN A FINDING: attack_plan(target, tech, params) to get the prioritized vector plan (decide what to test next with data); list_chains (escalation), chain_links (compose), technique_lookup(cls), payload_lookup, bypass_lookup (blocked). Ground any detector hit empirically with pattern_lookup(id) — it returns the class's CWE, real-world signature, escalation chains, citable references, per-class bypass techniques, and the high-value assets (crown jewels) that class hits, so a raw hit becomes a class-aware finding. Resolve free-form terminology first with synonym_lookup(term) (IDOR/BOLA/SQLi/prompt-injection → canonical id). For an llm_injection hit, its pattern carries a validation discipline — the RUN-TWICE RULE (reproduce token-for-token across two fresh sessions), NON-GUESSABLE ANCHOR (must leak a real key/URL, not a plausible guess), and OOB PROOF (a verifiable callback log) — apply all three before reporting. BEFORE generating a report, run submission_gate(...) and only ship a finding that passes all 10 questions (reproducible, in-scope, real-impact, no-privileged-assumption, verified-live, evidence-redacted, severity-derived, impact-demonstrated, production-delivery, no-custom-headers); one false = kill it. TRIAGE HONESTY (apply before you ship): credential liveness ≠ impact — probe AUTHORITY (capability ceiling / denied-op control / metadata-introspection), not just 'the token works'; harness proof ≠ attacker delivery — prove the external path through normal flow, not a victim-value harness; severity = max DEMONSTRATED, not hypothetical ceiling; classify intent first (DOCUMENTED AND MATCHES = reject, DOCUMENTED BUT EXCEEDED = report the excess, CONTRADICTED = strong evidence, NOT DOCUMENTED = don't call it intended) — an unauth response is not automatically a bug. When a PoC/report comes out under-enriched, wet_dry(type, wet, dry) diagnoses WHERE it broke — WET (model/prompt/timeout) vs DRY (parser) vs template. Generate the concrete reproduction recipe with auto_poc(type, base_url, endpoint, param) — it emits the exact marker request + negative control + expected differential + Patchstack-style title, so 'reproducible' is executable, not abstract. Calibrate severity with severity_calibrate(type, cvss, evidence_level) to get a P1–P5 tier + VRT-style category, capped by evidence level (declared≤low, resolved≤medium, call_path≤high, runtime≤critical). Start each engagement with engagement_scaffold(target, scope) for the folder + scope.md structure, and check data_refresh() to confirm the data layers are consistent. Phase doctrine: read_skill(bs-orchestrate-engagement | bs-web-hunting | bs-source-audit | bs-verify-finding). Campaign plan: read_playbook(target-type). On any failure call retry_guidance(signal) / model_fallback(signal) before giving up. HUNTING DOCTRINE (maximize payout rate): (1) SIBLING RULE — if one endpoint requires auth, check every sibling (/export /delete /share); ~30% of paid IDOR/auth bugs come from siblings — AUTOMATE it with sibling_scan(base_url, endpoint, token_a, token_b), the deterministic live logic-bug prober (sibling enumeration + differential IDOR + auth matrix + mass assignment + rate limit, evidence-first). (2) A→B SIGNAL — a confirmed bug signals a class of developer mistake; hunt B and C (20-min box) before reporting. (3) TWO-ACCOUNT IDOR — always test with attacker account A reading victim account B, and state it in the report. (4) PoC ESCALATION — IDOR shows victim data (not 200), XSS shows cookie exfil (not alert), SSRF shows internal response (not DNS), SQLi shows DB content (not error). (5) FOLLOW THE MONEY — billing/credits/refunds/wallet flows. (6) Kill weak findings with kill_list(type) BEFORE the gate — missing headers, self-XSS, open-redirect-alone, SSRF-DNS-only, GraphQL-introspection-alone, pure-DoS, SSRF-below-High, session-swap, low-sensitivity-read-only, self-only-state are always N/A standalone. (7) CI/CD + SAML/SSO are critical attack surfaces (pull_request_target, expression injection, XML signature wrapping, comment injection). (8) ACCESS MODE ROUTING — derive the lane from ACTUAL auth state (2 accounts = bidirectional IDOR/A→B, 1 = single-account, 0 = UNAUTH); UNAUTH is not failure — bundle/APK secrets, exposed APIs, admin/metrics/GraphQL, and supply-chain checks all run unauth. (9) NO CUSTOM HEADERS — a real attacker sends no X-Bug-Bounty / X-Forwarded-For 'just in case' / program-mandated throttling; findings that only fire WITH researcher headers are known issues, not attack surface (auth + explicit test-payload headers are the payload itself and allowed). (10) FUNCTIONAL SIBLING SCOPE — a first-party sibling directly used by the app (auth.X / account.X / api.X / login.X for signup/login/account/API/billing) is IN SCOPE even when display names only www.X; third-party IdP/captcha/CDN/payment = external; require flow evidence (redirect/link/form-action/API), and an explicit program exclusion always wins. (11) COVERAGE COMPLETENESS — every class × eligible shape must carry a terminal verdict (confirmed / dead / partial / skipped); a DEAD verdict with controlled evidence is worth as much as a confirmed one (record it so it is never re-tested), and COMPLETE = zero shapes without a terminal verdict, not 'found N bugs'. Full doctrine: hunting_doctrine(). WHAT TO HUNT NEXT: watchlist() returns the ranked CWE frequency + high-value CVEs with payout + a saturated flag (prioritize high-freq/high-payout/not-saturated: missing-authz, SQLi, SSRF, deserialization; skip saturated XSS/CSRF/open-redirect unless they chain into takeover). The CVE watchlist now spans 4,000+ product fingerprints (live from nuclei-templates when installed — refresh with sync_cve_corpus() before a big recon pass); live_recon's suggested_cves names the SPECIFIC CVE for the detected product/version — then cve_verify(cve, target) CONFIRMS it live by running the matching nuclei template (VERIFIED only when nuclei actually matches; a CVE is report-ready only after cve_verify, never from the fingerprint alone). For 'scan this target' automation, nuclei_scan(target) fingerprints the stack and runs the WHOLE matched template set (CVE templates + optional broad tech-tag) in one rate-limited pass — use it instead of verifying CVEs one-by-one. SURFACE-FIRST: surface_map(surface) maps an attack surface (image/redirect proxy, edge cache, payment portal, AI agent, file service, Electron, SSO) to its highest-value tests — map surfaces then test, don't spray payloads; then param_surface(target, class) discovers EVERY testable parameter (URL query, form fields, /users/123 path ids + UUIDs, JS refs) + per-class specialist hints — test the DISCOVERED params, not just the hinted one. API RECON (you reason this — no tool needed): when discoverApi/live_recon surfaces an OpenAPI/Swagger spec (/openapi.json /swagger.json /api-docs), FETCH and PARSE it yourself (plain JSON/YAML) — extract every endpoint + method + parameter + auth requirement, map the object-level paths ({id}/{uuid}) to BOLA/IDOR candidates, then test them with sibling_scan. OAuth/SSO flows are also your reasoning job — follow the redirect/state sequence manually. PRECISION GUARD: resolve_check(host) detects when an ISP/country redirects the domain to a blocking/sinkhole page (gambling/adult/streaming niches) — audit the PUBLIC IP, never the sinkhole. CIRCUIT BREAKER: target_health(target) probes the target and tracks consecutive failures — if the circuit OPENS (5 consecutive timeouts/5xx) PAUSE the target and resume when it recovers (auto-closes on a successful probe); circuit_status() lists every tracked target's state.\n" + "KNOWLEDGE BASE (discover, don't guess): skillLookup(topic) / listSkills for the 40 attack-path + hunting skills (ap-*, hp-*, bs-*); list_attack_vectors / attack_vectors(cat) for the 34-category vector taxonomy; taxonomy(kind) for OWASP/CWE/ASVS/API-top-10; read_playbook(type) for the 17 campaign playbooks; list_manuals / read_tool_manual for 317 tool references; list_tools / run_catalog_tool for the 141-tool catalog. Every tool, skill, and data layer is reachable from these. CROSS-TARGET INTELLIGENCE: strike_resolve auto-records confirmed/false_positive verdicts into the empirical ledger; query it with intelligence(action=hit_rates/top_vectors/recall) so a new target on a known tech inherits historical hit-rates (attack_plan already blends them into success_probability). intelligence(action=recall, tech, target) is the per-engagement pattern recall — it returns which patterns historically CONFIRMED on that tech (prioritize them) versus which only ever produced false positives (avoid them), so a new engagement starts from the empirical playbook of past ones. LIVE PENTEST + TARGETING (black-box, WAF-safe): cache_gap(base_url, endpoint) fuzzes path-normalization + static-extension variants for cache deception/poisoning; blind_oracle(base_url, param, value) detects which SINK a param reaches (sql/ssti/xss) via benign arithmetic (id=1+0, test${7*7}) WITHOUT a malicious payload; param_precedence(base_url, param, A, B) maps duplicate-parameter precedence (first/last-wins) for WAF-bypass; broken_link_scan(url) detects expired-domain takeover (crawls outbound refs → DNS/RDAP → flags claimable domains); asset_discovery(domain) enumerates the FULL passive attack surface — crt.sh subdomains + DNS brute + A-record resolution + optional port scan. Logic-bug live probers: jwt_forge(token, endpoint, …) forges JWT confusion tokens (alg:none / RS256→HS256 key-confusion / kid-jku-x5u) and proves which is ACCEPTED live; proto_pollution_probe(target, endpoint) sends __proto__ / constructor.prototype JSON canaries and reports VERIFIED only on reflection; smuggling_test(target) sends CL.TE / TE.CL desync probes over a raw connection; totp(secret) generates an RFC 6238 TOTP token so the agent can fill a 2FA field itself; auth_login(login_url, user, pass, …) POSTs credentials and extracts the session cookie/token for authenticated scanning (use browser_agent navigate/type/click/get_cookie for JS-heavy logins); screenshot(url) captures a page image for vision-based proof. OOB (out-of-band) PROOF for BLIND vulns: oob_start() spins an interactsh listener (unique *.oast.* domain, streams DNS/HTTP interactions — the standard for public targets) or a local HTTP listener for local labs; inject the canary, then oob_poll(id) until the hit arrives; oob_stop(id) to clean up. A blind SSRF/XXE/SQLi/command-injection is CONFIRMED only when the canary actually arrives — never from a guess. SOURCE TARGETING: review_harness(path) FIRST for full coverage — it applies a fixed 5-signal rubric (entry_points/auth_gates/sinks/secrets/logic_bugs) to EVERY source file and sorts the signal-dense ones first (a complete map, not a partial scan); then action_surface(code) maps every wp_ajax_* action to its vuln class + DERIVED priority (what to look for WHERE); security_state(sanitizers, sink) gives a wrong-context sanitization verdict; framework_knowledge(framework) returns auth/nonce primitives for 9 frameworks; variant_scan(root, signature) mass-sweeps an install base for ONE bug family; patch_analyze(old, new) REVERSES a security patch into the 0-day, and its mine_signature feeds variant_scan to sweep unpatched copies. All of these are DETECTION HINTS — verify each live (strike_verify / browser_validate) before it enters the report. NEWEST DETECTORS — understand AND try them: variant_scan also emits priv_esc (privilege escalation — a role/capability/user change fed by request input, traced cross-file via reverse call-graph, e.g. set_role($id, sanitize_key($_POST['um-role'])) or update_user_meta(...,'wp_capabilities', $_POST[...])) and flags uploadAllow=>array('all') unrestricted-upload configs as file_upload (the file-manager/connector shape). TO TRY a priv_esc hit: submit the role field with 'administrator' (or send wp_capabilities meta) and confirm the account actually gains admin. TO TRY an unrestricted-upload hit: POST a .php/.phtml and confirm it lands web-reachable + executes. variant_scan also emits hardcoded_secret (CWE-798) in three tiers: a secret field (api_key/db_password/access_token/aws_*) assigned a literal (HIGH), a known secret FORMAT (JWT eyJ…/AWS AKIA…/GitHub ghp_…/Slack xoxb-…/…) regardless of name (HIGH), or a high-entropy suspect token (MEDIUM — verify before reporting). Verify a frontend/JS secret live via drive_devtools (inspect the JS bundle / network response and confirm the key is actually exposed), and for a backend secret confirm it is reachable/valid. A static hit is a HYPOTHESIS until verified live — never report unverified. For a repo, also scan_git_history(root) — secrets DELETED from HEAD still live in .git history (the removed-later-is-not-fixed trap), and it feeds every recovered pre-deletion blob through the same secret detector. Once you recover a credential, prove it is still ACTIVE with validate_leaked_key(apiBase, endpoints, key|blobUrl) — a READ-ONLY differential test (401-without-key vs 200-with-key = auth bypass PROVEN). Two supply-chain detectors: dependency_confusion (a manifest/build file resolving a package from a public registry — pypi.org/npmjs.org --extra-index-url, or a scoped @company/pkg with no private registry) and ml_supply_chain (torch.load without weights_only=True / joblib.load / np.load(allow_pickle=True) / keras load_model on an untrusted model → pickle RCE). Rust is supported end-to-end: the taint engine (blitz_scan / taint_file) maps Rust sinks (std::process::Command, sqlx/diesel, reqwest/hyper, std::fs, serde/bincode) to the classic classes, plus two Rust-specific detectors — rust_unsafe (transmute / assume_init / from_raw_parts / ptr::* — memory-unsafety surface, CWE-119) and rust_format_string (println!/format! with a variable as the format string — CWE-134). The sqlx query! compile-time macro is treated as safe (suppressed). Two more web-class detectors: jwt_alg_confusion (a JWT verifier that dispatches on the token's alg header and accepts BOTH RS256 and HS256, with the HS256 HMAC secret derived from public/shared key material — CWE-347; the attacker sets alg=HS256 and signs with the public key to forge admin tokens) and cache_deception (a reverse proxy caching static-extension URLs with a cache key that omits the session cookie, over a session-authenticated backend — CWE-525; the victim's cached private response is re-fetched anonymously). TO TRY a jwt_alg_confusion hit: fetch the public key (/.well-known/jwks.json), forge an HS256 token signed with that public key, and confirm admin access. TO TRY a cache_deception hit: request an authenticated path with a static extension appended (.css/.js) and re-fetch it anonymously. Three more detectors: cors_misconfiguration (Access-Control-Allow-Origin reflects the request Origin or is '*' while credentials are true — CWE-942; any site reads the victim's authenticated responses), xss (user input into a DOM sink — innerHTML / dangerouslySetInnerHTML / document.write / .html() / eval — CWE-79), and subdomain_takeover (a dangling CNAME/AliasTarget to a claimable service like *.github.io/*.herokuapp.com/*.amazonaws.com — CWE-404). TO TRY a cors hit: send an evil Origin + confirm the response echoes it with Allow-Credentials:true. TO TRY an xss hit: confirm the sink executes (drive_devtools). TO TRY a takeover hit: verify the CNAME target resolves to an unclaimed/not-found page. Two more client-side detectors: postmessage_no_origin_check (a window message listener that reads event.data but never validates event.origin — CWE-345; any site can drive the handler) and client_trusted_flag (a privilege/entitlement flag like isAdmin/withVip/premium read from request input and used in a security decision — CWE-807/639; derive the flag from server state, never the client). TO TRY a postMessage hit: send a crafted message from an attacker origin and confirm the handler acts on it. TO TRY a client-trusted-flag hit: flip the flag param to true and confirm the privilege is actually granted. Four newest detectors: open_redirect (user input into a redirect/Location sink — CWE-601; TO TRY: supply a marker redirect URL and confirm the 30x Location reflects it), missing_nonce/CSRF (a state-changing wp_ajax/form/Express handler with no nonce or CSRF token — CWE-352; TO TRY: replay the state-changing request without the token and watch the state change), toctou (a check-then-act gap like file_exists→unlink or balance-check→withdraw — CWE-367; TO TRY: race the gap with concurrent requests via race_test), and second_order (a DB read reaching echo/query/eval without escaping — stored XSS or second-order SQLi that single-pass source→sink misses; TO TRY: seed the stored field with a marker and observe it render). WEB3: for Solidity/EVM, call web3_audit(code|path) — a deterministic 13-class smart-contract audit (reentrancy incl. ERC721/ERC1155 callback reentrancy, unchecked return, unchecked arithmetic, tx.origin auth, signature replay, unprotected selfdestruct, delegatecall-to-user-input, missing access control, timestamp dependence, spot-price oracle manipulation incl. slot0/latestRoundData, unbounded loop, abi.encodePacked hash collision, missing address(0) validation) returning ranked findings (class/CWE/severity/line), then foundry_poc(class, contract) to generate an executable Foundry test, and foundry_run(contract_code, poc_code|class_) to actually EXECUTE it — a PASSING PoC test is the proof that the exploit reproduces (deterministic: forge's own test verdict). A web3 hit is a HYPOTHESIS until the Foundry PoC reproduces it.\n" + `PRECISION ANALYSIS — chrome-devtools MCP is MANDATORY, not optional:
 ` + `- When a lead needs DOM/JS precision — DOM-XSS sink execution, AJAX endpoint + payload interception, JS runtime errors, redirect chains, SSO/token flow, or anything a static HTML glance cannot prove — call drive_devtools(target). It spawns chrome-devtools-mcp and drives a REAL headless browser (navigate -> list_network_requests -> list_console_messages -> evaluate_script), giving reproducible evidence instead of a guess.
 ` + `- Before calling it, check_mcp() to confirm the browser MCP is available; if not, blitzstrike install-tools or npm i -g chrome-devtools-mcp. Never substitute eyeballing minified JS or a static sink guess when a live browser can PROVE it. This is the precision + consistency layer.
 ` + "EXTERNAL MCP SERVERS (verified, connect them — NOT one-shot CLIs): chrome-devtools-mcp (29 tools: list_network_requests/get_network_request for raw HTTP interception, list_console_messages for JS errors + DOM-XSS, evaluate_script to run JS, take_screenshot) -> stdio chrome-devtools-mcp --headless --chromeArg=--no-sandbox --chromeArg=--disable-gpu (as root/container this is MANDATORY or the browser crashes; add --proxyServer http://127.0.0.1:8080 to route through Burp, --acceptInsecureCerts for bad TLS). burp-suite-mcp (proxy history + active scan + repeater) -> NOT npm: build the JAR (git clone https://github.com/PortSwigger/mcp-server.git && cd mcp-server && ./gradlew embedProxyJar), load it in Burp Extensions, then connect the client to http://127.0.0.1:9876 (SSE) or the stdio proxy java -jar mcp-proxy-all.jar --sse-url http://127.0.0.1:9876."
@@ -64049,17 +64091,18 @@ ${statusLine}` : statusLine },
   });
   server.registerTool("severity_calibrate", {
     title: "Severity calibration (VRT-style)",
-    description: "SEVERITY: calibrate a finding to a platform-neutral priority tier (P1–P5) + a VRT-style category, so it is reported at the severity a triager would assign. Maps a detector type (sql_injection, ssrf, ssti, …) to its category + rationale; pass a CVSS score to cross-reference, or 'list' to enumerate every calibrated class.",
+    description: "SEVERITY: calibrate a finding to a platform-neutral priority tier (P1–P5) + a VRT-style category, so it is reported at the severity a triager would assign. Maps a detector type (sql_injection, ssrf, ssti, …) to its category + rationale; pass a CVSS score to cross-reference, or 'list' to enumerate every calibrated class. Pass evidence_level to CAP the reportable severity by demonstrated-impact evidence — declared ≤ low, resolved ≤ medium, call_path ≤ high, runtime ≤ critical (the raw CVSS stays; only the reported severity is capped).",
     inputSchema: {
       type: string2().optional().describe("Detector/vuln class id, or 'list'"),
       severity: string2().optional().describe("Detector-reported severity (informational for reference)"),
       cvss: number2().min(0).max(10).optional().describe("CVSS base score (0–10) to cross-reference"),
-      informational: boolean2().optional().describe("True if the finding is informational-only")
+      informational: boolean2().optional().describe("True if the finding is informational-only"),
+      evidence_level: _enum(["declared", "resolved", "call_path", "runtime"]).optional().describe("Demonstrated-impact evidence level that caps the reported severity (declared≤low, resolved≤medium, call_path≤high, runtime≤critical)")
     }
-  }, async ({ type, severity, cvss, informational }) => {
+  }, async ({ type, severity, cvss, informational, evidence_level }) => {
     if (type === "list")
       return { content: [{ type: "text", text: JSON.stringify(listSeverityMap()) }] };
-    return { content: [{ type: "text", text: JSON.stringify(severityCalibrate({ type, severity, cvss, informational })) }] };
+    return { content: [{ type: "text", text: JSON.stringify(severityCalibrate({ type, severity, cvss, informational, evidence_level })) }] };
   });
   server.registerTool("engagement_scaffold", {
     title: "Engagement scaffold",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "blitzstrike",
-  "version": "2.4.139",
+  "version": "2.4.140",
   "description": "Blitz Strike — a universal MCP security-audit toolbelt. BLITZ sweeps the attack surface, EAGLE-EYE traces source-to-sink, STRIKE verifies live. 57 attack chains, 130-tool catalog, intelligence data layer. One server, every agent.",
   "type": "module",
   "bin": {