Overview
- Anthropic reports its new Claude Opus 4.6 autonomously identified more than 500 previously unknown, high‑severity vulnerabilities across widely used libraries including Ghostscript, OpenSC, and CGIF.
- Frontier Red Team testing placed the model in a virtualized environment with debuggers and fuzzers to assess out‑of‑the‑box behavior, and Anthropic says every finding was validated and triaged with emphasis on memory‑corruption bugs.
- Some issues have already been patched by maintainers, including a CGIF heap buffer overflow fixed in version 0.5.1 that the company says required conceptual understanding of the LZW algorithm and GIF format.
- To curb misuse, Anthropic is rolling out internal monitoring “probes” to flag risky activity in real time and expanding enforcement tools with the ability to block malicious traffic, while acknowledging added friction for legitimate researchers.
- OpenAI is framing similar capabilities cautiously, rating its new GPT‑5.3‑Codex model high for cybersecurity risk and tightening rollout with delayed full API access, restrictions on high‑risk use cases, and a vetted trusted‑access program.