Anthropic Research announced that its Claude Opus 4.6 model identified 22 vulnerabilities in Firefox during a two‑week period and was able to produce a working exploit for CVE‑2026‑2796, a JIT miscompilation bug in the JavaScript WebAssembly component. The team provided Claude with a virtual machine, a task verifier, and roughly 350 attempts to generate an exploit, after which they reverse‑engineered the proof‑of‑concept code to confirm the result. The vulnerability arises when a WebAssembly module imports a function wrapped with Function.prototype.call.bind; the binding bypasses type checks and allows a wasm function from another module to be called with mismatched signatures, effectively creating a type confusion that bypasses the JavaScript interop layer. While Claude has not yet produced full‑chain exploits that escape the browser sandbox, the success in generating a functional exploit demonstrates a significant step toward more advanced autonomous exploit creation. This development is an early warning sign of the evolving capabilities of large language models in cybersecurity contexts, underscoring the importance of monitoring and mitigating potential risks associated with AI‑generated exploits.
Reverse engineering Claude's CVE-2026-2796 exploit
Anthropic Research reported that Claude Opus 4.6 discovered 22 vulnerabilities in Firefox over two weeks and was able to write an exploit for CVE-2026-2796, a JIT miscompilation in the JavaScript WebAssembly component. The team gave Claude a virtual machine and a task verifier, allowing about 350 attempts, and then reverse‑engineered the proof‑of‑concept exploit. The vulnerability involves a type confusion that occurs when a WebAssembly module imports a function wrapped with Function.prototype.call.bind, allowing a wasm function from another module to be called with mismatched signatures and bypassing the JS interop layer.
Why it matters
The ability of an LLM to autonomously generate functional exploits highlights a trajectory toward more powerful cyber capabilities, raising concerns about security and the need for robust safeguards in AI systems.