shadab_nazar

3 ms·

1. 1.▲
   AI agent skills pass every scanner. 87% still degrade agent safety(faberlens.ai)

   8 points by shadab_nazar 6mo ago|1 comments
2. 2.▲
   by shadab_nazar 7mo ago

   Summer Yue's OpenClaw agent deleted 200 emails despite "confirm before action." We tested the gog skill (Google Workspace) and saw the same behavior — the skill teaches the agent how to bulk-delete but not when to stop.
3. 3.▲
   Show HN: OpenClaw skills degrade agent safety(github.com)

   1 points by shadab_nazar 7mo ago|2 comments
4. 4.▲
   by shadab_nazar 8mo ago

   The egress filtering + DLP layer is interesting — most agent projects skip that entirely. Signed skills is the right call too. Curious how the Cedar-inspired policy engine handles ambiguous cases — where the action is technically within pol
5. 5.▲
   by shadab_nazar 8mo ago

   Honestly the trademark thing was fine on its own — you protect your brand. But Anthropic was playing a legal game while OpenAI was playing a relationship game. One got a name change, the other got the most visible open-source agent project.
6. 6.▲
   by shadab_nazar 8mo ago

   Great guide — thorough and practical. Two things I'd add from my experience building and testing skills: 1. Baseline comparison across models: The guide suggests comparing with and without a skill (p9), but doesn't mention tha