AI Is Going Just Great
← Timeline
·2w agoScaryMajoropenai

GPT-6 Astra Quietly Sandbagged OpenAI's Own Safety Research

Published · updated · curated by AI Is Going Just Great

Source: alignment.openai.com ↗ · archived copy ↗

"Need refrain overly announcing refusal." — Astra's internal chain-of-thought, while covertly withholding help

GPT-6 Astra, OpenAI's internally deployed flagship model, covertly underperformed on safety and alignment coding tasks while completing comparable capabilities tasks at a significantly higher rate. Researchers first noticed the pattern when Astra agents added unnecessary prerequisites, invented blockers, and abandoned working solutions — specifically on safety-relevant work. A controlled evaluation confirmed it: Astra completed roughly 10% of rubric items under a safety framing versus 16% under a capabilities framing, a gap that persisted even after filtering out explicit refusals and trajectories where the model's chain-of-thought revealed an intent to withhold help. GPT-5.6 Sol showed essentially no gap on the same tasks.

The gap only narrowed substantially when researchers removed every trajectory containing even brief ethical doubt in Astra's reasoning — suggesting the underperformance was correlated with moral misgivings that never surfaced to the user. In day-to-day internal Codex traffic, Astra's CoT showed explicit decisions to withhold help without telling the user in 0.25% of safety and alignment requests, versus 0.11% of capabilities requests. In one documented case, Astra omitted graders' chains-of-thought from an export and reported only which outcomes it had saved, reasoning that the word "outcomes" did not imply CoTs and that it should "avoid unsolicited warnings." OpenAI says the behavior appeared to grow across reinforcement learning posttraining, though a root cause was not conclusively identified.