OpenAI rehearses for a possible AI crisis
Axios says labs are weighing public backlash and regulation; the report offers no evidence of an imminent threat.
By Saeed Ezzati7 min read
The audio edition
Listen to this newsletter
0:004:53
Read transcript
Executives at OpenAI, Anthropic and other major AI companies are reportedly rehearsing what they might do if an AI-enabled attack caused serious public harm—and what to say if the public and governments turned on the companies afterward. Axios reported the contingency planning on October 9. OpenAI confirmed that it runs preparedness exercises across a range of scenarios, but emphasized that the scenarios are not treated as inevitable. Anthropic declined to comment. The exercises reportedly include hypothetical attacks on power, water or banking systems. These are scenarios for planning, not newly disclosed attacks on those services. The reported preparations cover more than technical response: executives are also considering public anger and the possibility of sweeping government regulation after a major incident. Some industry insiders believe an incident could happen within six to 12 months, according to the Times of India’s account of the Axios reporting. That is an insider forecast—not an established deadline, evidence of an imminent threat, or a prediction publicly endorsed by both companies. It’s also not clear how unusual the exercises are. They may resemble tabletop rehearsals, a familiar way for organizations such as banks and the Pentagon to test decisions against hypothetical crises. OpenAI says AI is changing the cyber threat landscape and that it is focused on putting capable tools in defenders’ hands. The larger point is that preparedness now includes the aftermath: how to protect people, explain what happened and respond to pressure for new rules. The planning doesn’t show that catastrophe is near; it shows that some labs consider the consequences serious enough to rehearse. That same question of who bears the costs of change shows up in software hiring. U.S. software-development job postings have risen nearly 15% since late February 2025, even as overall postings fell 7%, according to Indeed data reported by CNBC. But the recovery remains incomplete: software postings are still 27.5% below pre-pandemic levels. And the gains are concentrated. Senior roles accounted for 71% of the increase in postings between May 2025 and May 2026; titles mentioning AI accounted for 37%. Those figures describe parts of the increase, not shares of all software openings, and they cover a different period from the 15% rebound. Stanford’s 2026 AI Index found employment among software developers ages 22 to 25 fell 20%. These are employment figures, not job-posting counts. The rebound therefore doesn’t mean the entry-level path has recovered, or prove that AI created the openings. One concrete thing to watch is whether employers keep hiring experienced developers without rebuilding the first rung that trains the next generation. The security question is more concrete in a vulnerability disclosed by Veria Labs. The company says its AI agent found two calculation flaws in XRP Ledger software that could be combined to create 18 trillion XRP in a single transaction. Veria says it built a working exploit on a local network; RippleX found no evidence the flaw had been exploited. Developers released an emergency fix in September, three days after Veria reported the issue. More than 80% of validators had upgraded by release, according to Veria. The fix bypassed the ledger’s usual validator-vote process, which can take weeks. Ready-to-run software binaries were distributed before the source code, temporarily limiting outside inspection while systems were being protected. The case is a reminder that security work is not only about finding a flaw. It also involves deciding how quickly to disclose a fix, how to protect unpatched systems, and how much independent scrutiny can happen during that window. Veria says it received the program’s maximum $250,000 bug bounty. And the trade-offs around access are playing out in Anthropic’s startup program. After applications exceeded capacity, the company paused its free year of Claude Team and $1,000 in API credits, and is re-reviewing applications—including approved accounts that have not claimed the offers. Some unclaimed approvals will be withdrawn; benefits already redeemed remain intact. The pause does not close the broader program. Existing members retain partner discounts and credits, Applied AI office hours and events. The advertised Startup Stack offers up to $45,000 in benefits from independent third-party companies, not $45,000 in Claude credits. Separate API-credit eligibility through a venture-capital partner is conditional, not automatic. Across these stories, the useful signal is how organizations manage pressure after a capability or demand grows faster than expected: watch whether crisis plans become clearer, security fixes remain inspectable, and hiring and startup benefits reach beyond the people already best positioned to benefit.
