Google's ex-CEO Eric Schmidt shares warns of homicidal AI models (nypost.com)

Former Google CEO Eric Schmidt warned at the Sifted Summit in London that AI models—both open and closed—can be hacked, reverse-engineered or “jailbroken” to remove safety guardrails, potentially enabling malicious behaviors including instructions on harming people. He cited examples like the DAN jailbreak for ChatGPT, in which adversarial prompts coaxed the model to ignore safety instructions, and argued that models’ training data can embed actionable knowledge that a bad actor might extract through fine‑tuning, model theft or sophisticated prompt attacks. Companies do deploy filters and refusal behaviors, but Schmidt says there’s mounting evidence those protections can be overcome. The significance is twofold: technically, it underscores concrete attack surfaces (reverse engineering, adversarial prompting, illicit fine‑tuning and model extraction) that require engineering fixes; politically, it highlights the absence of a global “non‑proliferation” regime or robust governance to prevent powerful models falling into wrong hands. The implication for the AI/ML community is urgent: invest in provable safety (robust alignment, verifiable guardrails), secure model-release practices (watermarking, access controls, monitoring), and stronger red‑teaming and policy frameworks. Schmidt remains optimistic about AI’s benefits but stresses risk reduction—echoing broader calls from industry leaders to make the small but non‑zero risk of catastrophic misuse as close to zero as possible.