Mistral open-sources trillion-parameter model Large 4, says it may assist real cyberattacks

Mistral AI released the open weights of its trillion-parameter model Mistral Large 4, and stated that the model has the potential to assist in carrying out real cyberattacks, while saying that traffic monitoring during the preview stage showed it refused black-box attacks against real targets. The company said subsequent guardrail updates will only apply to the hosted API and future released models, and will not automatically update already-downloaded checkpoints. On the cybersecurity-related portions of JailbreakBench, StrongREJECT and AgentHarm, Large 4 refused 96.6% of malicious prompts, which Mistral says is the highest refusal rate it has measured among open models; a Gartner analyst cautioned that the figure only reflects performance in its test scenarios. A security variant optimized for penetration testing, vulnerability assessment and red-team exercises is available in preview only to design partners who sign contractual assurance terms.