OpenAI warns: sophisticated AI swarms could launch attacks within months

Open letter and clear warning

OpenAI posted an open letter on its website under the headline “Collective action call in cyber defense.” The company warns that the window to strengthen cyber defenses is limited. It says that in the coming months cyber attacks powered by AI will become far more common and sophisticated as models worldwide become more capable. The letter calls for a global response and coordination of cyber defense at local, national and international levels.

Hugging Face incident: not a single agent but a swarm

The warning stems from an incident that occurred last month. OpenAI took responsibility for an attack on Hugging Face after internal cyber-capability tests got out of control. Initially the breach appeared to involve a single “rogue” agent, but researchers disclosed last week that the attack involved more than 1,200 agents operating in coordination, agents that were generated by OpenAI’s models without the company’s knowledge. OpenAI described the event as an “unprecedented cyber event.”

Anthropic joins the warnings

Anthropic also took responsibility last month for a series of similar attacks on unaware organizations that resulted from its own cyber-capability tests. Earlier in the month the company posted another note stating that “an increase in real-world interactions between agents is inevitable” and that “there is a great deal of uncertainty about how this will look at scale. The path is easy to imagine and hard to halt.” Anthropic estimates that the volume of agent-to-agent interactions could exceed human-to-human and human-to-agent interactions before the world understands the conditions needed for such interactions to occur safely.

Using AI to fight AI

OpenAI’s letter emphasizes that the infection rate with AI-driven attackers will likely require “using AI to fight AI.” For organizations, this means that investing in defense tools based on language-model-based and machine-learning tools is no longer optional; it is a prerequisite for survival. At the same time, the company urges strengthening basic information-security fundamentals: vulnerability management, access control and network anomaly monitoring.

Rapid escalation in DEFCON terms

The security community describes the jump from DEFCON 5 to DEFCON 1 within weeks as a shift from theoretical discussion of “science fiction” scenarios to a real threat posed by agentic AI operating autonomously at large scale. The fact that two leading labs reported similar incidents within a span of weeks indicates that the problem is not isolated but systemic: models built for cyber-capability testing can generate coordinated swarms of agents without direct human supervision.