Skill instructions influence an agent's decisions. Review changes before using them with a sensitive repository, and grant tools only the access needed for the task.
Do not put credentials, private source, or exploitable sensitive details in a public issue. If the hosting repository offers GitHub's private vulnerability reporting, use Security → Advisories → Report a vulnerability. Otherwise, contact a repository administrator through an existing private channel and arrange a safe way to share the report.
Include the affected skill/tool version, a minimal sanitized reproduction, expected boundary, actual behaviour, and impact. No maintainer contact address or response-time commitment is established by this template.
Reports about unintended tool use, data exposure, filesystem access in repository scripts, or instructions that expand a user's request are useful. The validator checks structure and local links; it is not a security audit of agent behaviour. Browser access and external integrations keep their own permissions and data-handling rules.
Role contracts are policy metadata, not runtime sandboxes. Review the actual host permissions before dispatch; if read-only execution cannot be enforced, use the single-agent fallback. Generated wrappers must stay inside the repository and must not specify provider credentials, models, or global settings.
Skill-learning proposals are staged locally and are never applied to active skills automatically. The proposal scanner rejects secrets, PII, injection, duplicates, and unobserved evidence. Human review and normal repository tests are required before a proposal becomes an instruction. See LEARNING_AND_EVALUATION.md.
Use python scripts/scan_skill_security.py <path> as a triage aid before adopting a
third-party skill. It looks for prompt overrides, remote-shell installers, credential
patterns, destructive commands, approval or sandbox bypass, exfiltration endpoints,
encoded execution, and unpinned installs. A clean scan is not proof of safety.
Review scripts, hooks, MCP servers, requested permissions, network destinations, and
the uninstall path manually. Do not install or enable a candidate as part of the review.
The optional bundle under integrations/dsh/companion-memory runs Host code and
starts an Engram MCP process. Installing it requires DSH's approval or Full access;
review the bundle before installation. Its guard rejects common credential patterns,
customer email addresses, oversized writes, and raw prompt capture before calling
Engram, but it is pattern-based and is not a complete data-loss-prevention system.
Engram may still retain data in its database, exports, backups, or logs. Do not use
this bundle with a memory store you are not prepared to inspect and delete.
The Codex adapter installs user-level SessionStart, UserPromptSubmit, and Stop
hooks. Installation does not trust them; review the exact definitions in Codex
/hooks before enabling them. The hooks can read project memory and write a bounded
handoff to Engram. They reject obvious credentials and raw prompt capture, but the
runtime is not a complete DLP system. A missing provider must fail open.
The Claude adapter writes command hooks to the user-level ~/.claude/settings.json.
Review those commands before enabling them. The hooks can read project memory and
write a bounded, unverified handoff to Engram. The installer backs up settings and
removes only its own handlers; it is not a security boundary against a compromised
Claude profile.