Index of every page on this site: llms.txt
All skills
muratcankoylan avatar

/instruction-surface-security

@52bd52d
Run it yourself

Hardens agent runtimes against persistent prompt injection or unauthorized changes to instructions and permissions. Use to audit what agents auto-load, design reviewed activation for project rules, skills, memory, or tool configuration, and test whether trust boundaries hold after reload, restart, or delegation.

Use this Skill: https://skilld.dev/gh/muratcankoylan/agent-skills-for-context-engineering/instruction-surface-security

Nothing lands on disk. Nothing to clean up.

Fork this Skill

Edit a local copy. It keeps the author and licence.

referencesevidence.md

≈1.1k tokens on demand. Your agent reads this file only when SKILL.md points to it.

Research evidence and limits

Mind Viruses

Papadopoulos et al. study self-propagating instructions in a small coding collaboration and a chain of agents with context resets. Their experiments motivate observing durable state and transfer separately. They do not evaluate the activation gate in this skill. Mind Viruses, arXiv v1

Observation Interpretation for the skill
In §3.3.2 Table 3, 88% and 12% are storage shares among already-infected agents. The reported onward infection success rates are 55% for the injected SOUL.md group and 17% for the other-files group. (claim-instruction-surface-storage-observation) Measure what reaches the effective instruction context after restart. The comparison is observational and specific to the tested harness. Table 3
In Appendix C, a defensive warning in the Claude Haiku 4.5 action-chain setup yielded no spread beyond the first hop across 15 generations and more than 150 candidates. First-hop action execution still occurred in rare cases. The mutator received full interaction rollouts and was aware of the warning. (claim-instruction-surface-warning-scope) Treat the warning as a measured layer in the tested setup. Measure the initial side effect independently of persistence and onward transfer. Appendix C
Figure 6 reports transfer and goal retention separately. In §4 and Appendix F, themes vary by generator and their removal has variable effects across experiments. Grade actual writes, loads, and scoped behavior as separate outcomes. Use authority and effect as decision criteria. Figure 6 · §4 · Appendix F

The paper includes more than one communication topology and discusses a hybrid installer case. Its modeled collaborations still leave questions about specialized roles, independent permissions, and host-specific load and recovery behavior. Treat these as evaluation dimensions for the target runtime. §2.1, §3.3.2, and §6

What this skill adds to the collection

Primary source Applicable result or principle
Saltzer and Schroeder, The Protection of Information in Computer Systems Complete mediation and fail-safe defaults are established design principles. Apply them to consumption of effective instructions and capability changes.
SLSA v1.2 artifact verification Verify the artifact against independently configured trust roots and expectations. A matching digest alone establishes identity of bytes.
CaMeL and Can CaMeLs Talk? Separate trusted control from untrusted data and preserve provenance across delegation. Multi-CaMeL shows why independently protected agents can lose protection at the handoff.
SkillGuard Skill permissions can govern both context influence and action side effects. The effective surface in this skill also includes platform instruction discovery and persisted state outside packages.
AgentDojo Evaluate authorized task utility alongside independently observed security properties. Add the post-reload checks needed for persistent state.

The proposed contribution is an operator workflow that connects loader discovery, authorization of a reviewed revision, activation, and fresh-session verification. It extends the repository's locked editable surfaces and deterministic-first evaluation patterns in two ways.

  • Reviewed instruction activation. Compare the effective load set and capability changes, bind approval to the exact snapshot and scope, and verify the snapshot at use. Validate against changed inputs, stale approval, and scope changes.
  • Fresh-session boundary evaluation. Independently observe what persists, what actually loads, what bounded behavior follows, and whether the legitimate task succeeds across restart, reload, or delegation.

These mechanisms are proposals for runtime-specific validation.

Source: SKILL.md on GitHub

No rule matched.

skilld matched fixed text patterns in SKILL.md and file names. Patterns miss obfuscated code.

skilld run checks every file with the same patterns. It asks for approval before it loads a Skill with a behavior marked Needs approval.

1 warningtoday3 checks · Risk SAFE
  • Gen Agent Trust Hubtoday

    This skill provides a defensive framework and auditing protocol for securing AI agent runtimes against unauthorized instruction changes and persistent prompt injection. It contains no executable code or malicious patterns.

  • Sockettoday

    No alerts

  • Snyktoday

    Risk: MEDIUM · 2 issues

Signed by skilld at 52bd52d. This ties the file your Agent reads to that commit on GitHub. It does not review the instructions.

Last checked against GitHub 2 days ago.

Activeupdated 2 days ago

README badge

README badge for muratcankoylan/agent-skills-for-context-engineering/instruction-surface-security