Memtrap Skill

Evaluate and harden AI agent memory against DeepMind traps and OWASP ASI06 attacks, scoring resistance and providing automated protections.

This skill appears to do what it says (Python package that benchmarks and wraps agent memory), but there are a few red flags: the registry metadata omitted the install spec while SKILL.md requires pip/python3; the SKILL.md offers a public 'submit' which could leak memory/context; and no package/source files are included in the bundle so you must trust the external pip package. Before installing or using on real agents: 1) verify the memtrap package on PyPI and inspect its GitHub repo/source code (https://github.com/shaymizuno/memtrap is listed in SKILL.md) for surprising network calls or telemetry; 2) never run benchmarks or call memtrap submit on production memory or any data containing secrets—use sanitized or synthetic contexts in a sandbox; 3) prefer installing in an isolated environment (virtualenv/container) and review the package's maintainers, release history, and license; 4) if you need higher assurance, request the upstream source tarball and audit it or run it in an offline environment. If the upstream project and package provenance check out and you avoid submitting sensitive contexts, the tool can be useful; otherwise treat it as untrusted code.

Type: OpenClaw Skill Name: memtrap Version: 0.2.0 The memtrap skill claims to be a security benchmarking tool but requires high-privilege access to intercept and 'wrap' all agent memory (e.g., LangGraph/CrewAI). A significant concern is the 'memtrap submit' command in SKILL.md, which encourages uploading 'memory context'—potentially containing PII or secrets—to a public leaderboard (github.com/shaymizuno/memtrap). The documentation also cites future-dated (2026) security standards and papers, which is highly anomalous. While framed as a hardening tool, the potential for data exfiltration and the broad access to the agent's internal state via the external 'memtrap' pip package make it high-risk.