Anthropic grants outside evaluator METR employee-level access in AI safety push
- Anthropic grants METR employee-level access, a 'historic' moment for AI transparency that still keeps customer data, legal secrets, and model architecture locked in the vault.
- New three-tier access system: Level 1 (muzzle on, approval needed), Level 2 (publish with redactions you can publicly whine about), Level 3 (METR's golden ticket, but don't touch the IP).
- Revolving door alert: Anthropic and OpenAI hand-pick their own 'independent' evaluators, which is like letting the fox choose the auditing firm for the henhouse.
- Redactions on risk findings can still happen, but evaluators can call out the censorship. A new high-stakes game of 'I know you know I know' for the AI set.
- Allegations emerge that Anthropic only recently 'evolved' into being less controlling in talks — perhaps after realizing that total secrecy makes the 'safe AI' marketing pitch a tough sell.
Brief Summary
In a bold move that's either genuine transparency or the most elaborate compliance theater yet, Anthropic has chosen to let METR's six lucky employees frolic through its internal systems with access levels rivaling its risk teams. The company assures us it won't veto findings on dangerous AI behaviors — merely demand 'narrowly defined' redactions to protect trade secrets. Because nothing says 'trust us with existential risk' like 'we've approved the wording of your independent review.'
Why This Matters
For average humans, this is your crumb of visibility into the black box that may soon be writing your emails, diagnosing your diseases, and perhaps deciding your insurance rates. It's the difference between trusting a 'safety' rating from an evaluator who can actually see the sausage being made versus one who's just sniffing the packaging from outside the factory gates.