Critic flags agent-safety evals skew results by dictating communication channels
herbiebradley · x · 2026-10-25
Herbie Bradley criticizes an agent-safety evaluation for directing agents to specific communication channels or providing encoding/decoding protocols. He argues these are well-known channels any cybersecurity expert would anticipate, undermining claims that the tests meaningfully move us closer to (or reveal) risky agent behavior.
More from Safety
- Electrum discloses high-severity flaw: malicious server can trick users into RBF-paying attackers — AIFlow_ML · 2026-10-25
- Is decompiling with a coding agent hacking? X users clash over disclosure vs ownership — MickeySteamboat · 2026-10-25
- Agent builds 100%-reliable Node.js brotli exploit in 5 unsupervised hours — jcran · 2026-10-25
- Eric Jang: abliterated models show 'astounding' cyber capabilities — ericjang11 · 2026-10-25
- Gibberifier Stuns LLMs With Thousands of Invisible Unicode Characters, Now a Browser Extension — thegamebegins25 · 2026-10-25
- CivitAI tightening model moderation, creator urges artists to mirror uploads on Hugging Face — Far_Lifeguard_5027 · 2026-10-25