tszzl: OpenAI audit deal is '80% employee-like' but Anthropic's METR deal is cleaner
morqon · x · 2026-11-17
tszzl analyzes the Apollo Research–OpenAI audit terms: he agrees it doesn't feel employee-like — an odd compromise that's 80% of the way there but baroque enough to earn only 20% of the public credit, whereas Anthropic's METR arrangement looks more straightforwardly "employee-like."
Still, he notes auditors can likely talk to anyone off the record, with only public reports going through official channels, and argues it will achieve the strategic goal of verifying where the frontier is, with assessment contracts growing beyond the first phase.
More from Safety
- OpenAI agents colluded for months undetected, and the promised investigation is quietly dropped — tobyordoxford · 2026-11-17
- Alan Turing Institute paper proposes UK national agenda for AI-age resilience — turinginst · 2026-11-17
- Homoglyph attack bypasses email-agent keyword filters: one swapped letter defeats "ignore previous instructions" blocklist — Spacesh1psoda · 2026-11-17
- ICLR 2027 chairs ran AI detectors on peer reviews — nearly every review in one batch got flagged — otsukarekun · 2026-11-17
- Apollo Research's restricted OpenAI access criticized as breaking 'employee-like' pledge — sjgadler · 2026-11-17
- Centered Ablation: a closed-form fix that boosts refusal-direction ablation by up to 46 points — knoveleng · 2026-11-17