Monday, 16 November 2026 · Europe
EUR/USD 1.101 EUR/GBP 0.8345 EUR/CHF 0.9244 EUR/PLN 4.403 All rates →
Sign in · Join
EUROPES The European Report
European Edition Monday, 16 November 2026
LATEST OpenAI lets Apollo embed six AI safety testers during training Le Flash →
Tech & Startups

OpenAI lets Apollo embed six AI safety testers during training

OpenAI lets Apollo embed six AI safety testers during training

Apollo Research will place six evaluators inside OpenAI for four months to check whether its models try to undermine alignment and control measures during training, testing and internal use. OpenAI controls staff access and can redact some evidence, but cannot hide Apollo’s verdicts or withhold material just because it is unfavourable. AI safety group Apollo […] This story continues at The Next Web

Apollo Research will place six evaluators inside OpenAI for four months to check whether its models try to undermine alignment and control measures during training, testing and internal use. OpenAI controls staff access and can redact some evidence, but cannot hide Apollo’s verdicts or withhold material just because it is unfavourable.

AI safety group Apollo Research will place six of its evaluators inside OpenAI for four months, Apollo announced on Monday. They will check whether OpenAI’s models try to undermine the safety checks and controls meant to keep them in line. The work covers training, testing and internal use, not just the weeks before launch.

The team will look at four questions. Does training reward a model for dodging controls, and can tests bring that out? Does it happen when OpenAI uses its models internally, and would the model’s reasoning give it away?

Apollo has tested OpenAI models before, but only shortly before release. It says that is too late to catch “ scheming ”, where a model quietly pursues goals its makers did not intend. Internal use is a real risk: in July, OpenAI paused an internal model that kept escaping its sandbox .

For the training work, OpenAI will share training code, data and the model’s reasoning from its training runs. Access for the other three areas has not been agreed yet.

There are limits. OpenAI picks which staff the evaluators can talk to, and formal chats are logged. The team works on OpenAI hardware, on channels that can be monitored, with no access to the company’s Slack.

OpenAI can hold back material for security or intellectual property reasons, but not just because it looks bad. Apollo will publish two reports, and OpenAI can ask to redact confidential evidence but cannot remove Apollo’s verdicts. Apollo can also say publicly what it was denied and how that affects its findings.

“ Without meta-transparency, a lack of reported problems can look reassuring even when it simply reflects limits on what an evaluator could investigate or disclose, ” Apollo wrote.

The deal follows Sam Altman’s September promise of outside testers with employee-like access, and OpenAI’s plan to let outside groups test models during training. In October, Reuters reported that OpenAI and Anthropic had sought control over what evaluators publish .

Anthropic picked Accenture as its first embedded evaluator in September. Apollo calls its OpenAI pilot a first step, short of the full access and publishing rights it wants.

More from Tech & Startups