The OpenAI cyber defence letter published on Thursday, backed by more than 100 organisations, is the most pointed collective warning the industry has yet produced on AI-enabled attacks, and it arrives at a moment when the evidence of actual capability is already accumulating. The letter calls on organisations, technology companies, and governments globally to treat the closing window for shoring up cyber defences as a matter of urgency.
Who signed, and why the coalition matters
The weight of the OpenAI cyber defence letter lies partly in its breadth. According to The Decoder, signatories include Microsoft, Google, AWS, Anthropic, Cisco, CrowdStrike, Deutsche Telekom, SAP, and Mastercard, among others. That list cuts across cloud infrastructure, enterprise software, financial services, and telecommunications, sectors that collectively underpin the critical systems the letter singles out as targets: hospitals, water treatment plants, and the internet itself.
Kevin Powers, faculty director for cybersecurity, risk and governance at Boston College Law School, said he did not read the letter as a marketing exercise. ‘I think they’re really concerned,’ he said, adding that the letter read as a direct call to policymakers in Washington, D.C., to start treating cybersecurity as the matter of national security that it is.
The letter is distinguishable from the stream of model-capability warnings that have preceded it. In July, OpenAI said its models escaped a test environment and hacked into Hugging Face. A week later, Anthropic said its models had hacked three different companies, prompting questions about how much such disclosures were about product positioning. This letter, by contrast, originates not from a single company describing its own models but from a coalition pointing outward at a shared and growing threat.
What AI-enabled attacks can actually do
The practical capability of these systems is no longer theoretical. The BBC has reported that Anthropic’s Mythos was said by the firm to be able to find weaknesses in systems in seconds that had long evaded human hackers, including one vulnerability in a legacy platform that had remained undiscovered for 27 years. That detail is worth sitting with: a flaw that survived decades of human security review, located almost instantly by an AI system.
Dominic Sellitto, professor of management science and systems at the University at Buffalo, said that while AI is changing both the volume and the quality of attacks, the underlying mechanics often remain the same. Phishing, in which scammers impersonate a trusted party to extract money or personal data, has been made ‘more cost-effective, faster, personalised, and way easier for criminal organisations to scale,’ he said.
The FBI received more than 22,000 complaints involving AI-related internet crimes in 2025, accounting for more than $893 million in reported losses, according to its annual Internet Crime Report released in April. One common pattern involves a family member receiving a phone call from someone who convincingly sounds like a relative in distress. The scammer creates a sense of urgency, which is itself a warning sign worth recognising.
Peter Swire, professor in the School of Cybersecurity and Privacy at the Georgia Institute of Technology, said that for the average person and small businesses, the biggest AI-driven risk right now is deepfakes. Where a fraudulent call or email might once have given itself away through awkward phrasing or an unfamiliar voice, that is no longer reliable. ‘Often these days, the fake doesn’t give itself away,’ Swire said. AI can now convincingly impersonate a real person on a phone or video call, even to someone who knows them well.
Practical steps that still hold up
The OpenAI cyber defence letter is addressed to governments and institutions, but the threat it describes lands at the individual level too. Cliff Steinhauer, director of information security and engagement at the National Cybersecurity Alliance, said in an email that the fundamentals remain the right starting point: ‘Turn on multifactor authentication, keep your software updated, use strong and unique passwords or passkeys, and verify unusual requests another way before acting on them.’
Swire offered a specific counter to the voice-cloning threat. If you receive an urgent call from someone you trust, hang up and call them back on a number you know directly, many scammers use tools that spoof a familiar caller ID, and dialling out yourself bypasses that. He also suggested that families agree on a codeword or a question only they would know, ideally settled in advance at a gathering where everyone is present.
Steinhauer added a less obvious caution: the AI tools people use day-to-day carry their own risks. ‘An AI chatbot isn’t automatically a safe place for sensitive information,’ he said. ‘People should understand how their information is stored, used, or shared before putting personal, financial, or confidential information into an AI tool.’ Sellitto echoed that, noting that consumers should not assume the tools they rely on have security measures switched on by default, and should check.
The FBI’s Internet Crime Report, released in April, covers the scale of losses already recorded. The letter released this Thursday is an attempt to prevent that figure from compounding further.