Cyber Filters Bite, and the Open-Closed Cyber Gap Widens

@MiaAI_lab reported that OpenAI’s Codex refused to review a plan written by Anthropic’s Opus 5.5 because it flagged the plan as a cybersecurity risk. “Even Opus was surprised,” she wrote.

She then posted what she called proof: the block came from OpenAI’s cybersecurity content filter, with a message reading in part, “This content was flagged for possible cybersecurity risk. If this seems wrong, try rephrasing your request… apply for Daybreak access…”

Separately, @kimmonismus noted that the gap between Chinese open-weight models and US closed models appears to be widening again on cybersecurity benchmarks. Epoch AI’s Cyber ECI showed roughly a 6-point gap in early 2025 and 11 points now, with GLM-5.3 at 160 against 171 for the US closed frontier. Open models are improving, he wrote, but in this domain the leaders have pulled further ahead.