The AI outfit Anthropic debuted Claude Opus 4.6 and it is already being flogged as a weapon in the security arms race.
Anthropic deployed Opus 4.6 to a sandboxed environment and tested whether it could detect bugs in open-source code without being given hints. The team gave it access to Python and vulnerability analysis tools, including classic debuggers and fuzzers, but no instructions or specialised knowledge.
The model still identified more than 500 previously unknown zero-day vulnerabilities using its out-of-the-box capabilities, and each was validated by either an Anthropic team member or an external security researcher.
Anthropic, head of frontier red team Logan Graham told Axios: “It’s a race between defenders and attackers, and we want to put the tools in the hands of defenders as fast as possible. The models are extremely good at this, and we expect them to get much better still.”
The bugs ranged from flaws that could be exploited to crash a system to nastier issues that could corrupt memory, which is where defenders start sweating. In a blog post, Anthropic said Claude uncovered a flaw in GhostScript, a popular utility for processing PDF and PostScript files, that could cause it to crash.
Claude turned up buffer overflow flaws in OpenSC, a utility that processes smart card data, and CGIF, a tool that processes GIF files, which is the sort of unglamorous plumbing attackers love.
Anthropic thinks Opus 4.6 is a major win for security teams because open-source code underpins everything from enterprise software to critical infrastructure, and it is chronically under-audited.
Anthropic, head of frontier red team, Logan Graham, said: “I wouldn’t be surprised if this was a way in which open-source software was secured.”
In many cases, Claude used its upgraded reasoning to change course when traditional tools proved ineffective. For the GhostScript flaw, it dug through the project’s Git commit history after fuzzing and manual analysis failed to find anything useful, then probed the wider codebase to see if similar bugs were lurking.
For CGIF, Claude even wrote its own proof of concept to demonstrate the vulnerability was real, which is impressive and somewhat alarming in the same breath.
Anthropic says it has added new security controls to Claude Opus to spot and respond to adversaries trying to abuse these cyber capabilities. The blog post warns: “This will create friction for legitimate research and some defensive work, and we want to work with the security research community to find ways to address it as it arises,” while also proposing real-time detection that could block traffic it deems malicious.
