The model is an instrument in a human-led exploit chain
Thesis: This case shows a commercially available model participating in high-end vulnerability research, while leaving the human researcher responsible for directing and validating the work. Forbes reports that Xbow engineer Brendan Dolan-Gavitt used GPT-5.6 to chain three previously unpatched Chrome vulnerabilities. The report says Google confirmed the findings and awarded its maximum $250,000 bounty. Mechanism: The engineer prompted the model for ways to crack Chrome’s security and worked with it to generate the chain; the reported $40,000 in token spend is a cost of that collaboration, not a measured cost per validated vulnerability.
Why now: Commercially accessible models can now contribute to research on widely deployed software, making the boundary between defensive research and attacker capability increasingly consequential. This is a concrete example in a browser with multiple security layers, including sandboxes, rather than proof that models routinely bypass them. First implication: Security teams should treat model-assisted vulnerability discovery as a capability available to legitimate researchers and potential attackers, while keeping human review and responsible disclosure central.
Evidence boundary: The account is Forbes’ report, including statements attributed to Xbow and Google; it does not provide an independently reproducible evaluation of how much of the discovery or exploit chain came from the model. It also says two other Xbow exploits await Google review, so those claims should not be treated as confirmed findings. First watch condition: Watch for Google’s disposition of the two additional reported exploits and for technical disclosure that clarifies the model’s actual contribution.