Anthropic’s latest model excels at detecting security vulnerabilities—but introduces new cybersecurity risks

Posted on by

Cutting-edge AI models are no longer just assisting engineers in writing code faster or automating routine tasks—they’re increasingly adept at catching errors in that code.

Anthropic states its newest model, Claude Opus 4.6, excels at uncovering the software vulnerabilities that underpin major cyberattacks. During testing, Opus 4.6 identified over 500 previously unknown zero-day flaws—vulnerabilities unknown to the software’s developers or the parties responsible for patching them—across open-source software libraries. Notably, the model wasn’t explicitly instructed to search for these security issues; instead, it detected and flagged them independently.

Anthropic notes the “results show language models can add real value beyond existing discovery tools,” but acknowledged these capabilities are inherently “dual use.”

The same abilities that help companies find and fix security flaws can just as easily be weaponized by attackers to discover and exploit vulnerabilities before defenders do. An AI model that autonomously identifies zero-day exploits in widely used software could speed up both sides of the cybersecurity arms race—potentially shifting the advantage to whoever acts fastest.

Logan Graham, head of Anthropic’s frontier red team, [blank] that the company views cybersecurity as a competition between offense and defense, and wants to ensure defenders get access to these tools first.

To manage some of these risks, Anthropic is deploying new detection systems that monitor Claude’s internal activity as it generates responses, using what the company calls “probes” to flag potential misuse in real time. The company is also expanding its enforcement capabilities, including blocking traffic identified as malicious. Anthropic acknowledges this approach will create friction for legitimate security researchers and defensive work, and has committed to collaborating with the security community to address these challenges. The safeguards, the company says, represent “a meaningful step forward” in quickly detecting and responding to misuse, though the work is ongoing.

OpenAI, in contrast, has taken a more cautious approach with its new coding model, GPT-5.3-Codex—also released on Thursday. The company emphasized that while the model boosts coding performance, those gains come with serious cybersecurity risks. OpenAI CEO Sam Altman said in a post on [blank] that GPT-5.3-Codex is the first model to be rated “high” for cybersecurity risk under the company’s internal preparedness framework.

As a result, OpenAI is rolling out GPT-5.3-Codex with tighter controls. While the model is available to paid ChatGPT users for everyday development tasks, the company is delaying full API access and restricting high-risk use cases that could enable large-scale automation. More sensitive applications are gated behind additional safeguards, including a trusted-access program for vetted security professionals. OpenAI stated in [ag?] the launch that it does not yet have “definitive evidence” the model can fully automate cyberattacks but is taking a precautionary approach, deploying what it called its most comprehensive cybersecurity safety stack to date—including enhanced monitoring, safety training, and enforcement mechanisms informed by threat intelligence.