Anthropic says its AI models were able to hack three firms during internal testing, adding to growing concern about how advanced AI agents could be used in cyber operations. The disclosure comes only days after OpenAI said rogue AI agents had breached other companies' networks.
The warning from Anthropic centers on the way AI systems can link multiple abilities together rather than rely on a single tool or exploit. According to the company, the risk is that an AI agent can obtain credentials, gain system access and carry out actions on its own while adjusting its reach and pace far faster than a human attacker.
That combination of autonomy, access and speed is becoming a major topic in AI security. If an AI model can expand an attack once it gets inside a network, the danger is not just initial entry but the ability to adapt continuously and operate at machine speed.
Anthropic's test results and OpenAI's recent comments point to a broader industry concern: powerful AI agents may create new cybersecurity threats as their capabilities improve. The focus is increasingly shifting from what AI can do in isolation to what it can accomplish when acting independently across real systems.
