Cantina Security released apex-flash-1 as a post-trained version of glm-5.3-flash, describing the release as a small run to refine its training process before a larger deployment. The firm, which ranks first on the 2026 HackerOne US business leaderboard and has earned $1 million in bounties from companies including Apple and Coinbase, has built models ranging from 27B to over 1T parameters. The open weights and an abliterated variant are currently available on HuggingFace.
Wednesday, October 21, 2026
Story
Independent Tester's 8.6% Score Contradicts Cantina's 66.7% AI Claim
AI14 days ago
An open-weights model developed for vulnerability research failed to generalize its capabilities during recent third-party testing on the CyberGym benchmark. A tester using the vanilla Claude Code CLI recorded an 8.6% pass rate for the apex-flash-1 model, contradicting the 66.7% pass rate asserted by developer Cantina Security. The evaluation identified infinite retry loops, tool-call confusion and context decoherence as primary factors in the model's failure to perform outside of the developer's internal testing harness.
Looking for more context or analysis?
Tilt chat provides reliable, tailored insights, alerts, and automations.