Cryptographer Matthew Green Stirs Debate After Testing AI Exploit Capability, Offers $500 Bet

On November 8, cryptographer Matthew Green got into a public dispute with several infosec practitioners, with the core disagreement being whether AI models can develop or launch novel exploits with minimal human intervention. Green cited his own experiment, which cost roughly 700 美元 in token fees: he had the model mine vulnerabilities across a large number of binaries with almost no human oversight, and said the results made him "very uncomfortable," leading him to believe the community underestimates models' capabilities in this area.

Confirmed

• Green says he spent about 700 美元 in token fees having a model hunt for vulnerabilities in a large number of binaries with nearly no human intervention, and the results unsettled him.
• He noted that almost no one, including many experts, has actually sat down like a novice and put open-source models to work, so the community doesn't know these models' capability ceiling; he also mentioned evaluations including Daybreak.
• Green made a public bet with MalwareTechBlog: he put up 500 美元 that even "a blatant amateur like me" could use a safety-removed (abliterated) GLM 5.3 to develop a pre-authentication exploit against a recent, known target.
• He also pushed back on the view that "exploits don't matter, existing crime rings are already mapped out or are just Russia, so nothing will change," saying he's willing to listen and even hopes it's true, but disagrees.

Not Yet Confirmed

• The concrete output of the model experiment (what vulnerabilities were found, how exploitable they are) wasn't detailed in the post.
• There's no follow-up yet on the specific terms and outcome of the 500 美元 bet.

Why It Matters

• The debate cuts straight to assessing the real offensive capability of open-source models once guardrails are removed: if even security experts lack systematic first-hand testing, the industry's judgment of model capability ceilings may be systematically low, which affects vulnerability disclosure, open-source model policy, and defense prioritization.

2026-11-08 ~ 2026-11-08 · 5 related posts

• Cryptographer Matthew Green says infosec is fighting over whether models can craft novel exploits — matthew_d_green · 2026-11-08
• [source] Security researcher spends $700 in tokens testing AI exploit discovery, unnerved by results — matthew_d_green · 2026-11-08
• Matthew Green on the infosec debate: can models develop remote exploits autonomously? — matthew_d_green · 2026-11-08
• [source] Matthew Green: Even experts haven't stress-tested what open models can actually do — matthew_d_green · 2026-11-08
• [source] Cryptographer Matthew Green bets $500 an abliterated GLM 5.3 can find a pre-auth exploit — matthew_d_green · 2026-11-08