Breaking

Anthropic and OpenAI Back Embedded AI Safety Evaluators as Independence Questions Mount

September 17, 2026 at 16:21 EDT

Policy & Regulation
Security
AI Agents

Anthropic and OpenAI are moving toward embedding outside safety evaluators inside their AI labs, a governance shift sparked by Dario Amodei's call to slow the pace of frontier AI development while safeguards catch up. The proposal has been welcomed by evaluators seeking deeper access, but researchers warn that access alone will not make them independent watchdogs.

September 2026 · AI safety

Embedded evaluators enter the AI safety debate

Anthropic and OpenAI are backing inside access for outside evaluators. The open question is whether access becomes independent oversight or company-controlled review.

labs publicly committed to embedded evaluators

parts in Amodei's pacing plan

6-12

months cited for potential internet-scale agent risk

Evaluation windows remain short

3 daysAstra review

6 daysincident review

Short review windows are one reason evaluators want deeper, ongoing access to training checkpoints and internal records.

The proposed oversight chain

Embedded accessDesks, tools, permissions and live conversations

Training evidenceCheckpoints, logs, transcripts and reward environments

Public findingsRisk levels, incidents and access limits

Why supporters care

Internal access could reveal when concerning behavior emerges during training, before a finished model learns to pass a narrow test.

Why critics hesitate

Independence depends on evaluator choice, contract terms, disclosure rights, time, scope and eventual legal backing.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.