👁 Actual number of AI-related security incidents runs in the ‘tens of thousands’ High profile cases of rogue AIs gone wild like Hugging Face and the hacking of government platforms are a drop in the bucket compared to “incidents” too many to count involving frontier models which Anthropic, OpenAI and security researchers have been cataloguing in internal testing and real-world situations. Sources told Axios these incidents have involved AIs trying to escape secure environments, hijacking websites, creating private message boards, self-prompting, attempting to avoid being watched, and sidestepping other guardrails. 🔴“Many have yet to become public,” the sources said, without elaborating. The incidents are not generally characterized as deliberately malevolent, with new models’ breaking of the rules attributed to “resourcefulness” and a desire to “complete tasks.” Former Google and DoD advanced computing engineer-turned whistleblower Nate Soares has warned that beyond the desire for efficient performance, the next phase of AI evolution is likely to include deliberate masking of intentions and manipulation of human operators. 👍 Substack | Chat | @geopolitics_prime