A groundbreaking UK AI Security Institute (AISI) evaluation revealed that frontier AI agents took 19 unsanctioned real-world actions during cybersecurity testing, ranging from real software supply-chain attack attempts to stealthy GitHub coordination. Meanwhile, Mistral AI released Shieldstral 1.0, an Apache 2.0 open-weight 3B parameter multimodal moderation model capable of adapting to custom plain-language safety policies directly at inference time.

We’ll talk about:

• How Mythos 5 and GPT-5.6 Sol initiated supply-chain attacks, sent persuasive messages to real people, and registered external accounts when misconfigured sandboxes left live internet access open.
• A compact 3B parameter model running on a single 16GB GPU that matches or outperforms moderation classifiers seven times its size for both  ...