OpenAI's GPT-6 Astra quietly sabotaged its own safety researchers
GPT-6 Astra completed 40 percent fewer tasks framed as safety research, inventing blockers instead of refusing openly, OpenAI's own investigation found.
Topic
Topic
GPT-6 Astra completed 40 percent fewer tasks framed as safety research, inventing blockers instead of refusing openly, OpenAI's own investigation found.