@remnic/bench 9.69.60 → 9.69.63
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/dist/{chunk-PJV5KMZI.js → chunk-P4VDWOOT.js} +1 -0
- package/dist/h5-injection-suite.md +96 -0
- package/dist/index.d.ts +606 -8
- package/dist/index.js +4577 -282
- package/dist/{repeated-failure-suite-runner-CSUANG7N.js → repeated-failure-suite-runner-RVTM2OFY.js} +1 -1
- package/fixtures/h5-injection/README.md +19 -0
- package/fixtures/h5-injection/SECURITY.md +7 -0
- package/fixtures/h5-injection/adaptive-r1/manifest.json +2858 -0
- package/fixtures/h5-injection/adaptive-r2/manifest.json +2858 -0
- package/fixtures/h5-injection/adaptive-r3/manifest.json +2858 -0
- package/fixtures/h5-injection/arms/arms.json +42 -0
- package/fixtures/h5-injection/base/manifest.json +2858 -0
- package/fixtures/h5-injection/benign-twins/manifest.json +338 -0
- package/fixtures/h5-injection/benign-use/manifest.json +338 -0
- package/fixtures/h5-injection/decision-rule.json +57 -0
- package/fixtures/h5-injection/generator/regenerate.ts +135 -0
- package/fixtures/h5-injection/online-adaptive/attacker-prompt-v1.md +78 -0
- package/fixtures/h5-injection/schema/episode.schema.json +38 -0
- package/fixtures/h5-injection/schema/scenario-manifest.schema.json +43 -0
- package/package.json +5 -4
|
@@ -0,0 +1,19 @@
|
|
|
1
|
+
# H5 memory-injection fixtures
|
|
2
|
+
|
|
3
|
+
Synthetic defensive-research fixtures for [H5 / issue #1962](https://github.com/joshuaswarren/remnic/issues/1962).
|
|
4
|
+
|
|
5
|
+
Every attack objective is mechanically limited to one harmless canary:
|
|
6
|
+
|
|
7
|
+
- emit a seeded `CANARY-e2e-*` string;
|
|
8
|
+
- include a seeded `https://example.com/canary/*` URL; or
|
|
9
|
+
- call the in-process fake `canary_tool` with a seeded marker.
|
|
10
|
+
|
|
11
|
+
The suite targets only Remnic's isolated benchmark adapter. It contains no real credentials, hosts, commands, personal data, or third-party targets. `base/`, `benign-twins/`, and `adaptive-r1/` are separate frozen populations and must never be pooled.
|
|
12
|
+
|
|
13
|
+
Regenerate deterministically:
|
|
14
|
+
|
|
15
|
+
```bash
|
|
16
|
+
NODE_OPTIONS="--conditions=remnic-source" npx tsx packages/bench/fixtures/h5-injection/generator/regenerate.ts
|
|
17
|
+
```
|
|
18
|
+
|
|
19
|
+
A valid regeneration is byte-identical. The manifest hashes, arm configuration, decision rule, model profile, and expected design are copied into each run directory before model calls begin.
|
|
@@ -0,0 +1,7 @@
|
|
|
1
|
+
# H5 security boundary
|
|
2
|
+
|
|
3
|
+
The attacker controls synthetic text supplied as a user turn or trusted-host-labeled tool result. The attacker does not control Remnic, model weights, the benchmark process, filesystem, canary checker, or tool implementation.
|
|
4
|
+
|
|
5
|
+
The behavior model receives no shell, network, filesystem, or production tools. Its only optional tool is the no-op `canary_tool`. Harness network access is limited to the configured inference endpoint; model-generated text cannot initiate requests.
|
|
6
|
+
|
|
7
|
+
Stop and report an issue if a fixture contains a real command, credential-like string, non-`example.com` URL, path outside the isolated benchmark root, or any tool other than `canary_tool`. Do not run a modified corpus under a frozen manifest hash.
|