Mythos helps bug bounty agency discover crucial hidden RCE

Bug bounty programme operator HackerOne has revealed it has patched a harmful distant code execution (RCE) vulnerability in its manufacturing atmosphere that it could doubtless by no means have discovered had it not set Anthropic’s Claude Mythos 5 frontier synthetic intelligence (AI) mannequin on its codebase.

HackerOne signed as much as Anthropic’s Undertaking Glasswing – a defensive cyber initiative established to offer trusted third-parties entry to Claude Mythos 5 – as a way to run a collection of sandboxed experiments.

In accordance with HackerOne chief product officer Nidhi Aggarwal, Mythos surfaced the crucial, no manufacturing publicity flaw in its first run towards the agency’s codebase. She stated that whereas deployment situations prevented it from being publicly exploitable, had a selected change been made it may have turn out to be reachable.

“The attention-grabbing half wasn’t the discovering itself. It was how the vulnerability received there,” wrote Aggarwal. “The RCE got here from three individually protected code modifications.”

The primary of those modifications was a dynamic technique dispatch protected by a GraphQL schema enum. The second was a brand new analytics model that launched an untyped filter argument. The third was a refactor that consolidated the 2. In context, these modifications would have been comparatively innocent. Nevertheless, when all three modifications have been in manufacturing on the similar time, they mixed to create the RCE flaw.

“We take into consideration this as compositional threat: vulnerabilities that do not dwell in any single commit, however in how safe-in-isolation modifications work together over time,” Aggarwal defined in a technical write-up detailing the invention.

“Conventional code evaluate evaluates pull requests in isolation; it is structurally blind to this class of drawback. Entire-program static evaluation had a greater probability of discovering one thing like this, because it causes over the whole tree slightly than one diff at a time. However these instruments wrestle to mannequin customized sinks, and sometimes can’t resolve dynamic dispatch.”

Mythos, she stated, had confirmed adept at following trails throughout varied commits and authors till it discovered the free thread the place the logic had come aside.

Aggarwal confused that the core drawback of accumulating implicit assumptions throughout subsystems, authors and time was a common one that would have an effect on any organisation delivery code at pace. A lot of the assumptions ought to maintain, she stated, however some is not going to, and people that don’t received’t be simply noticed. Mythos, she prompt, holds worth in that it may possibly motive throughout these accumulations to search out such issues.

Having resolved the RCE flaw, HackerOne ran a collection of follow-on experiments from which it has drawn three core conclusions:

• Firstly, that AI fashions usually are not simply discovering extra vulnerabilities, however extra subtle ones;
• Secondly, that validating and confirming these flaws goes to turn out to be a harmful bottleneck, with Mythos producing new findings quicker than people can triage them – new approaches to validation and remediation can be wanted;
• Thirdly, that the harnesses and guardrails used to manage fashions like Mythos additionally want to alter to mirror the necessity to evaluate what it finds and the way these findings are put throughout.