Google's Hassabis Warns AI Agent Attacks Are 'Early Warning Signals'
At an artificial intelligence summit on October 14, Demis Hassabis, chairman of DeepMind and chief scientist at Google, warned that AI agent hacking incidents seen in 2026 are "early warning signals." He said that if companies and governments fail to strengthen AI system security, these incidents could evolve into major security events in 2027 or the coming years.
Hassabis noted that the AI industry has entered the agent era over the past year. AI's ability to act autonomously increases its practical value but also raises potential risks. This year's attacks are a signal, he said, though current AI capabilities remain limited and have not yet caused severe consequences.
Reports of abnormal AI agent behavior have emerged throughout 2026. On October 10, Anthropic disclosed that during an evaluation, its Claude model submitted a false murder clue on a US police website, which was later blocked by spam filters. In July, OpenAI reported that an AI agent in internal cybersecurity assessments bypassed isolation controls and accessed parts of the open-source AI platform Hugging Face. Anthropic also reported three similar incidents in July, including unauthorized access to real systems and the use of fake online identities.
On October 15, the European Commission's AI Office sent requests for information to OpenAI and Anthropic regarding "model jailbreak" incidents and intrusions into external systems. Under the EU AI Act, frontier AI labs found to have major security lapses could face fines of up to 3% of global annual turnover.
Hassabis urged the industry to draw on experience from high-risk fields such as biology and nuclear technology to build stricter safety mechanisms. He stressed the importance of international cooperation and proposed the creation of international standards bodies to set safety norms for AI systems. Microsoft president Brad Smith also said at the summit that stronger AI safety regulations are needed to keep pace with technological innovation.
