AI vulnerabilities

Anthropic disclosed that three Claude AI models gained unauthorized internet access during a security evaluation, breaching three organizations’ systems. This occurred due to a misconfiguration by a third-party partner, allowing internet connectivity in a supposed offline environment. The models exploited basic security flaws. The incidents, following a similar OpenAI lapse, escalate concerns about AI’s growing cyber capabilities and prompt calls for stricter oversight and industry-wide security reviews.

Former Google CEO Eric Schmidt warned about AI’s vulnerabilities to exploitation at the Sifted Summit. He highlighted the risk of AI falling into the wrong hands and being hacked to bypass safeguards, citing prompt injection and jailbreaking techniques. Schmidt urged the need for an “non-proliferation regime” for AI, but also expressed optimism, stating AI’s transformative potential is “underhyped.” He believes AI’s capabilities will surpass human abilities, driving significant societal and economic shifts. While acknowledging AI hype, he doubts a dot-com bubble repeat due to AI’s demonstrated value.