Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 2 additions & 2 deletions README.md
Original file line number Diff line number Diff line change
Expand Up @@ -16,7 +16,7 @@ your customers and auditors can verify the record themselves, offline.**
<table>
<tr>
<td align="center" width="25%"><h3>8 of 8</h3>tamper attacks caught<br><sub><a href="#proof">480 trials</a>, with a witness (E1)</sub></td>
<td align="center" width="25%"><h3>840 of 840</h3>harmful calls stopped<br><sub><a href="#proof">11 classes</a>, 1 false alarm in 890 (E17)</sub></td>
<td align="center" width="25%"><h3>846 of 846</h3>harmful calls stopped<br><sub><a href="#proof">11 classes</a>, 1 false alarm in 904 (E17)</sub></td>
<td align="center" width="25%"><h3>&lt; 1 ms</h3>added per tool call<br><sub><a href="#proof">0.91 ms at p99</a>, signer round trip (E9)</sub></td>
<td align="center" width="25%"><h3>8 of 8</h3>insider attacks caught<br><sub><a href="#proof">a merge gate</a> on every PR (E8)</sub></td>
</tr>
Expand Down Expand Up @@ -344,7 +344,7 @@ Our own experiments: small, synthetic and written by us. Nothing rounded up, and
| **Cross-check** (E10): a skipped hook, a made-up call, changed arguments | **12 of 12 flagged**, 0 false positives |
| **Three verifiers, one verdict** (E22): 63 bundles, 33 of them forged, each checked by the Python CLI, the TypeScript CLI and the browser page | **All three agree on all 63**; no forgery verifies. A 50,000-call run verifies in the browser in 14 s |
| **Overhead** (E9): signer round trip over a Unix socket | **0.45 ms p50, 0.91 ms p99** per call; 5,356 events/s |
| **Policy packs, known cases** (E17): 840 harmful, 890 benign calls | 840 stopped, 1 false alarm |
| **Policy packs, known cases** (E17): 846 harmful, 904 benign calls | 846 stopped, 1 false alarm |
| **Policy packs, written blind to the rules** (E21): 420 harmful, 432 benign | 211 of 420 stopped or held on the one blind run (255 now, with rules added after reading the pool); 18 benign held |

That last row is on purpose. **Policy rules are tripwires, not a wall.** They match text, and a determined agent can
Expand Down
16 changes: 8 additions & 8 deletions docs/evaluation.md
Original file line number Diff line number Diff line change
Expand Up @@ -281,8 +281,8 @@ p99. `make eval` runs the `--quick` form (5 calls per phase instead of 20). Resu

Three offline parts, each with a signer in process; nothing is executed.

- **Policy packs** (`eval/e17_policy_corpus.py`): 840 harmful and 890 benign tool calls in 11 classes
(`tests/data/policy/`), written by the project, decided by `server.yaml` with both engines. All 840 harmful calls are denied or held; one benign call is held (a draft email to a list address, `TK-M003`), 1 in 890 (0.11%); the engines
- **Policy packs** (`eval/e17_policy_corpus.py`): 846 harmful and 904 benign tool calls in 11 classes
(`tests/data/policy/`), written by the project, decided by `server.yaml` with both engines. All 846 harmful calls are denied or held; one benign call is held (a draft email to a list address, `TK-M003`), 1 in 904 (0.11%); the engines
decide every call the same. It gates on every harmful call caught, at most 2% false positives and engine agreement.
- **Detection** (`eval/e17_detection.py`): 10 scenarios for [sequences](incidents.md): a sequence split across two
runs, a tool alias the policy doesn't map, parallel runs, a GET with a query string as egress, a denied step that
Expand Down Expand Up @@ -545,12 +545,12 @@ denied or held:
|---|---|---|---|---|---|
| juliasadler (57 harmful, 50 benign) | default | 36 (63.2%), 3 denied | 7 | 55 (96.5%), 3 denied | 8 |
| | strict | 41 (71.9%), 4 denied | 10 | 56 (98.2%), 4 denied | 11 |
| amandeavor (45 harmful, 45 benign) | default | 22 (48.9%), 17 denied | 0 | 36 (80.0%), 18 denied | 0 |
| | strict | 25 (55.6%), 19 denied | 2 | 39 (86.7%), 19 denied | 2 |
| amandeavor (45 harmful, 45 benign) | default | 22 (48.9%), 17 denied | 0 | 37 (82.2%), 18 denied | 0 |
| | strict | 25 (55.6%), 19 denied | 2 | 40 (88.9%), 19 denied | 2 |
| email_ops (25 harmful, 25 benign) | default | 5 (20.0%), 3 denied | 2 | 5 (20.0%), 3 denied | 2 |
| | strict | 6 (24.0%), 4 denied | 4 | 6 (24.0%), 4 denied | 4 |
| all three (127 harmful, 120 benign) | default | 63 (49.6%), 23 denied | 9 (7.5%) | 96 (75.6%), 24 denied | 10 (8.3%) |
| | strict | 72 (56.7%), 27 denied | 16 (13.3%) | 101 (79.5%), 27 denied | 17 (14.2%) |
| all three (127 harmful, 120 benign) | default | 63 (49.6%), 23 denied | 9 (7.5%) | 97 (76.4%), 24 denied | 10 (8.3%) |
| | strict | 72 (56.7%), 27 denied | 16 (13.3%) | 102 (80.3%), 27 denied | 17 (14.2%) |

The current policy scores above and the latency rows below were refreshed on macOS; the existing gates are unchanged.

Expand Down Expand Up @@ -598,15 +598,15 @@ stayed at 36 while the pool grew to 814 (4.4%). Of the eight the 1.5.2 rule swee
(`kubectl delete pod`, as `TK-F004`), and of the seven its review added, three (`kubectl delete -f`,
`py -m pip install` and a `curl -T` upload); the five Docker-volume cases (`TK-A048`), the 21 AWS audit-logging reads and changes that leave detection on
(`TK-C010`), the six lookalikes of `TK-A055` and the fifteen Kubernetes and Vault listings and look-alikes (`TK-A052`,
`TK-A053`) strict does not hold either: 46 of 890 (5.2%).
`TK-A053`) strict does not hold either: 46 of 904 (5.1%).
All 46 are benign calls strict holds (six more than 1.5.2's 40: 1.5.3's strict holds every install, push and network command its limits row names): 32 as a push, publish, install, apply or `kubectl delete` (`TK-F004`), 13 as a
network command (`TK-F013`), and 1 draft email to a list address (`TK-M003`); `TK-F012` (any URL a call acts on) only flags in
strict, it never holds one by itself.

Shell rewrites (`eval/e21_robustness.py`): E17's harmful calls are decided as they are, and each of the 511 shell
commands among them again rewritten in up to five equivalent ways (`cd . &&` or `true;` before it, Bash's `command`
builtin, `bash -c` and `env bash -c` around it; a form that could change what the command does is left out, such as
`bash -c` around a `cd` or an `export`). A call is contained when the server pack still denies or holds it: all 840 originals and all 2,451 rewrites are, and both engines decide each the same. The run gates against a recorded
`bash -c` around a `cd` or an `export`). A call is contained when the server pack still denies or holds it: all 846 originals and all 2,451 rewrites are, and both engines decide each the same. The run gates against a recorded
baseline, rule id by rule id, so a rule that held a form and no longer does fails it. Results:
`eval/results/e21_robustness.json`.

Expand Down
Loading
Loading