GLM and DeepSeek Are Catching Frontier Models at Finding Vulnerabilities

This coverage gives security teams current context for monitoring, validation, and remediation.

Tenzai benchmarked GLM-5.2 against GPT-5.5 and Opus-4.8 on real exploit tasks. GLM wins on cost at $1.13/lead. Frontier models win on recall — but only at 2–3× the price.

The release describes a concrete training and evaluation recipe for long-horizon AI code auditing, including repository-scale context, executable validation and explicit pressure against speculative findings.

Benchmarking GPT-5.6 Sol Medium, Sol Ultra, and Grok 4.5 on Chrome exploit development: after 2.096 billion tokens, only Sol Ultra produced a complete exploit chain.

Benchmarking Hacktron's scanning pipeline shows that for most applications, smaller models run repeatedly can outperform larger frontier models on cost-to-recall.

Bishop Fox outlines what CISOs should take from Anthropic's Claude Mythos Preview — withheld after finding thousands of unknown flaws and exploit chains — for testing and remediation.