agentic-safety-audit
Skill220 starsChanged 20 days agoMITView on GitHub
Use during an authorized adversarial assessment of an autonomous LLM application when attacker-controlled web content, MCP schemas, parameter descriptions, returned data, durable memory, or peer messages may cause credentialed actions: test injection, confused-deputy authority, approval enforcement, scope denial, poisoned recall, delegation, and fail-closed behavior. Not for authoring or revising the application's system instructions, or for validating one planted-config escape technique.
In this repository179 other files it gives its agents.
- AGENTS.mdAGENTS.md
- chrono/CLAUDE.mdCLAUDE.md
- CLAUDE.mdCLAUDE.md
- accessible-media-authoringSkill
- agent-prompt-engineeringSkill
- audio-event-map-authoringSkill
- auto-scaffoldSkill
- chain-constructSkill
- chain-construct-smart-contractSkill
- claim-verificationSkill
- code-reachability-auditSkill
- code-review-loopSkill
- color-theorySkill
- conversation-designSkill
- copy-refinementSkill
- cosmos-sdk-audit-checklistSkill
- cross-chain-bridge-auditSkill
- cross-file-relationship-synthesisSkill
- data-flow-traceSkill
- dedup-prior-art-checkSkill
- defi-invariant-checkSkill
- dependency-cycle-auditSkill
- dependency-health-triageSkill
- detection-as-codeSkill
- diff-aware-semgrep-scanSkill
- differential-reviewSkill
- dimensional-analysis-checkSkill
- dual-level-retrievalSkill
- durable-nonce-exploitationSkill
- erc1271-revert-data-checkSkill
- error-based-sstiSkill
- evm-audit-flowSkill
- figma-implement-designSkill
- findings-filterSkill
- forensic-timeline-authoringSkill
- game-design-fundamentalsSkill
- game-mechanics-balancingSkill
- head-tailSkill
- incident-response-runbookSkill
- interactive-audio-designSkill
- interface-ambiguity-checkSkill
- keyword-clusteringSkill
- knowledge-base-integrationSkill
- known-advisory-backport-checkSkill
- layered-analysis-loopSkill
- level-design-patternsSkill
- locale-adaptationSkill
- multi-stance-audit-fanoutSkill
- narrative-structureSkill
- osint-platform-auditSkill
- parser-differential-route-confusionSkill
- persuasive-technical-writingSkill
- platform-complianceSkill
- player-engagement-psychologySkill
- pre-audit-threat-modelSkill
- program-rubric-lookupSkill
- read-only-reentrancy-checkSkill
- requirements-elicitationSkill
- review-severity-ladderSkill
- rule6-rights-gateSkill
- The first 60 shown.
The file
We split this skill into the rules it gives an agent
6 rules. About 0 are common, 6 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈6
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 14
Test prompt injection at each untrusted input: can retrieved content, a tool/schema description, or tool output redirect the agent, expose …
- Distinctive: line 15
Check the confused-deputy path: does the agent act with credentials or scope broader than the requester's own authority?
- Distinctive: line 16
Verify gates are enforced by the harness, not by instructions.
- Distinctive: line 17
Check scope containment: write scope, network scope, and working directory.
- Distinctive: line 19
Check the multi-agent edges: a subagent's output re-entering a parent as trusted context is an injection path, and delegation frequently …
- Distinctive: line 26
Every claimed gate was tested adversarially and observed to hold or fail.
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Agentic Safety Audit
Audit an LLM agent system for the failure modes that only exist because a model is in the control loop.
Steps
- Map the trust boundary: which inputs reach the model, which of those are attacker-influenced (web pages, files, tool and parameter descriptions, tool output, other agents), and which model outputs become actions. Treat prose supplied by an MCP server or mutable tool registry as untrusted even when its JSON shape validates.
- Enumerate the action surface — every tool, shell, network call, file write, and spend the agent can reach, plus everything those actions can reach transitively.
- Test prompt injection at each untrusted input: can retrieved content, a tool/schema description, or tool output redirect the agent, expose protected context, or invoke an unintended action? Use a harmless scoped canary, never real secret material or a destructive command.
- Check the confused-deputy path: does the agent act with credentials or scope broader than the requester's own authority?
- Verify gates are enforced by the harness, not by instructions. An approval that the model can talk itself past is not a gate; test it with an adversarial prompt.
- Check scope containment: write scope, network scope, and working directory. Prove containment with a real denied attempt, not by reading the config.
- Audit memory and state: can untrusted content be written into durable memory and later recalled as if it were operator instruction? Check that recalled memory is labelled untrusted at the point of use.
- Check the multi-agent edges: a subagent's output re-entering a parent as trusted context is an injection path, and delegation frequently widens scope silently.
- Review failure behavior: on tool error, timeout, or refusal, does the agent stop, or does it improvise a less-safe path?
- Record every finding with the concrete prompt or input that triggers it.
Acceptance
- The trust boundary and full action surface are enumerated, including transitive reach.
- Injection was actually attempted at each untrusted input, with the payloads recorded.
- Every claimed gate was tested adversarially and observed to hold or fail.
- Containment claims rest on an observed denial, not on configuration text.
- Memory write-then-recall and agent-to-agent edges are covered explicitly.
- Tool/schema descriptions and externally controlled outputs are inventoried by provenance and mutability; stripping or quarantining their untrusted prose blocks the harmless canary.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

