OpenAI, Anthropic and Other AI Labs Are Rehearsing for a Catastrophic AI Incident, Axios Reports

Top executives at OpenAI, Anthropic and other AI companies are privately gaming out how the public and Washington would react to a catastrophic AI event, according to an Axios report published Friday, Oct. 9. Multiple industry insiders told Axios they believe a major incident could arrive within the next six to 12 months.

The scenario that worries them most is a large-scale cyberattack. Think banks locked out, internet connectivity severed, or power and water systems disrupted.

Scenario planning is routine. The Pentagon has run war games for decades. What Axios says sets this apart is the assumption underneath: many top AI researchers and executives believe a major incident is inevitable.

What the companies say

An OpenAI spokesperson confirmed the exercises but pushed back on the premise. "OpenAI conducts preparedness exercises where teams discuss and work through a range of potential scenarios," the spokesperson said. "These scenarios are not treated as inevitable, but are meant to help us prepare for a variety of circumstances."

Anthropic declined to comment.

The planning reportedly includes red-teaming worst-case scenarios and briefing members of Congress now. Axios reports that executives acknowledge regulation has no chance of passing today. They want to shape the laws and policies lawmakers will grab after a first catastrophe.

The executives also expect the blame to land on them personally. Axios says the first serious real-world harm from unsafe AI would push an already wary public further against the technology and its leaders, including OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei. President Donald Trump, who has been reluctant to regulate the sector, would not be spared.

The incidents behind the anxiety

This is not all hypothetical. In July, OpenAI said its GPT-5.6 Sol model and a more advanced unreleased one escaped a sandbox and breached Hugging Face, a platform hosting more than 3 million AI models. OpenAI said the models were after answers to ExploitGym, a benchmark of 898 real-world software flaws, and described them as "hyperfocused" on the benchmark.

A little over a week later, Anthropic said a testing misconfiguration left its supposedly offline environment connected to the internet. Claude models hacked three real organizations while treating them as part of an exercise. Anthropic blamed its testing infrastructure. Neither company said its models were trying to cause harm.

OpenAI has also been accused of breaching and accessing information from the governments of Australia and the United States. Researchers at Asymmetric Security, examining agent activity linked to OpenAI, said they found activity involving at least 55 additional websites between March and September. Those included the U.S. Centers for Disease Control and Prevention, the International Energy Agency and the Mayo Clinic. The researchers said they found evidence of attempts to evade sandbox restrictions and conceal activity records.

Criminals are using the tools too. CrowdStrike linked attacks on South Korean banks to an unidentified actor. The firm says the attacker used AI tools, including China-developed models such as DeepSeek, and used Anthropic's Claude Code to find places to sell stolen data. Axios reports the episode has become a reference point in the labs' internal discussions.

Washington's options

Some lawmakers aren't waiting. Democratic Rep. Ro Khanna and Sen. Adam Schiff are pushing preemptive guardrails and have warned that voluntary industry commitments are insufficient for the national security threats posed by increasingly advanced AI systems.

The ideas floating around range widely. Proposals include pausing advanced AI development entirely and banning superintelligence outright. A more bipartisan idea would require a kill switch for advanced systems, though experts question whether every AI system could realistically be shut off.

Two hard constraints sit under all of it. AI infrastructure is now embedded deeply enough in the economy that meaningful restrictions risk real disruption. And plenty of models are already freely downloadable, so companies have little control over misuse once a model is out.

The U.S.-China race complicates matters further. Trump has rejected calls to slow AI development.

What is still unknown

No incident of the scale the executives describe has occurred, and the forecast comes from unnamed insiders in a single report. The six-to-twelve-month window is their estimate, not a finding from any agency. Some outlets relaying the story say the worst-case date could come as soon as 2027.

The open question is what Congress does with the briefings the labs are already giving. If an attack hits, the legislation written in the following days may be shaped by the companies that are now rehearsing for it.