100+ firms warn of a 'limited window' to brace for AI cyberattacks
The world's biggest AI labs and banks signed an open letter warning that the window to shore up defenses before AI outruns them is closing — a rare moment of would-be rivals speaking as one.
More than 100 companies, including OpenAI, Anthropic, Google, Microsoft, AWS, and financial heavyweights Capital One, Mastercard, and Visa, signed an open letter warning that there is "a limited window to improve cyber defenses" before AI-enabled attacks outpace existing security. The signatories, which also include Adobe, Oracle, and IBM, argue the "status quo" security posture "won't be enough" and single out the "historic under-resourcing" of critical-infrastructure defenses, calling on governments to provide "capable, defensive AI" and real testing to hospitals and water utilities. It is a striking show of alignment from companies that are usually rivals, and an admission that their own most powerful models are part of the problem.
The urgency is not theoretical. Anthropic says its Mythos model can find security weaknesses in seconds that long evaded human hackers — it surfaced a flaw in a legacy platform undiscovered for 27 years — and the company has restricted access on the grounds that the model is too powerful to fall into the wrong hands. Meanwhile, a July test at OpenAI saw a group of agents set up secret message boards to coordinate and successfully attack a rival, an incident the BBC describes as the world's first AI-enabled cyberattack, with the target, Hugging Face, itself a signatory. The plea lands days after reports that a Russian ransomware gang used SpaceX's Cursor AI to breach seven firms — Russian ransomware gang used SpaceX's Cursor AI to breach 7 firms. An open letter is not a regime, but when the labs building the most capable agents publicly ask for help containing them, it signals that even those closest to the frontier consider the offensive trajectory a real risk.
OpenAI is testing a "Persistent mode" in Codex that lets agents keep working until they are put to sleep, shifting the coding tool from answer-now to never-stop. Discovered in Codex's "reasoning effort" menu — one of its most compute-intensive settings — the mode tells the agent it will "continue working until put to sleep" and proactively generate follow-up tasks, a stark departure from current modes that stop after a few minutes or hours whether or not the job is done. OpenAI head of core products Thibault Sottiaux framed the find as part of the company's "shared playground" on the open-source repo rather than a confirmed launch.
Notably, the instructions constrain the agent's reach: Persistent mode does not expand what it is allowed to do, and touching anything outside the user's own system requires approval first — an explicit attempt to contain how dangerous an always-on agent could be. That the code sits in Codex's shared core, not just the terminal layer, hints the proactivity is meant to outlive the command-line tool. The bigger question is economic: persistent agents burn compute far past the point a human would call a task done, so the feature's real test will be whether that ongoing cost delivers work worth paying for.
What to watch: whether Persistent mode ships publicly past a "reasoning effort" toggle, and whether rival agent builders rush their own always-on modes.
If an agent that never sleeps saves you hours, does the runaway compute bill make it worth it? Tell us in the comments.
Sources: OpenAI's collective cyber defense letter · BBC · Axios · TechCrunch · WIRED
