agent-trust-boundaries
Skill9 starsChanged 4 months agoLicence not identifiedView on GitHub
Design LLM and agentic systems so that retrieved content is data and never instruction — instruction authority comes from the channel, tool authority is granted by deterministic policy before the model runs, and the untrusted-content + private-data + external-egress trifecta is never unmediated. Use whenever an LLM reads something it did not write: RAG over documents, web browsing, email/ticket/PR triage, tool or MCP output, file contents, another agent's output, or user-supplied text that reaches a system prompt. Trigger on "prompt injection", "jailbreak", "the agent followed instructions in the document", "should the agent be allowed to", "give the agent access to", "MCP server", "tool permissions", "multi-agent", "sub-agent", "autonomous agent", "RAG security", or any design where a model's output causes a side effect. Sibling of llm-out-of-the-loop — that skill keeps the model out of the decision, this one keeps the attacker out of the model's authority.
In this repository91 other files it gives its agents.
- abductive-engineeringSkill
- acquisition-order-of-volatilitySkill
- alert-triage-economicsSkill
- assume-breach-modelingSkill
- atomic-state-mutationSkill
- attack-surface-triageSkill
- audit-before-patchSkill
- authorization-surface-mappingSkill
- beyond-the-fixSkill
- beyond-the-sinkSkill
- claim-provenance-disciplineSkill
- client-side-trust-boundarySkill
- cloud-architecture-selectionSkill
- cloud-control-plane-reasoningSkill
- codebase-health-assessmentSkill
- concurrency-reasoningSkill
- container-trust-boundarySkill
- containment-under-uncertaintySkill
- control-effectiveness-vs-existenceSkill
- credential-material-triageSkill
- crypto-misuse-reasoningSkill
- data-leakage-huntingSkill
- data-provenance-mappingSkill
- daubert-defensible-writingSkill
- debt-closure-disciplineSkill
- deception-signal-qualitySkill
- decision-record-disciplineSkill
- dependency-provenanceSkill
- destination-driven-constructionSkill
- detection-engineeringSkill
- deterministic-coreSkill
- diagnosing-bugsSkill
- discriminating-proofSkill
- distributed-write-atomicitySkill
- dont-fall-in-love-with-the-bugSkill
- dual-use-behavior-adjudicationSkill
- exploitability-triageSkill
- falsifiable-testingSkill
- finding-custodySkill
- forensic-logging-designSkill
- forensic-persistenceSkill
- frontend-accessibility-by-constructionSkill
- frontend-component-boundariesSkill
- frontend-performance-budgetsSkill
- frontend-testing-strategySkill
- git-disciplineSkill
- honest-degradationSkill
- honest-technical-oratorySkill
- hypothesis-driven-huntingSkill
- incident-timeline-reconstructionSkill
- intel-source-evaluationSkill
- invariant-huntingSkill
- irreversible-action-gateSkill
- known-defect-dispositionSkill
- language-selectionSkill
- live-contract-compatibilitySkill
- llm-out-of-the-loopSkill
- model-evaluation-disciplineSkill
- non-overwhelming-uxSkill
- oracle-driven-fuzzingSkill
- The first 60 shown.
The file
We split this skill into the rules it gives an agent
8 rules. About 0 are common, 8 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈8
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
The licence could not be identified — read it at the source. We count its rules but don't quote them. Read them on GitHub.
≈ “Common” is an estimate: at least 1 in 5 of the 350,495 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.
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.