More than a hundred technology, finance and infrastructure companies signed an open letter last Thursday warning that AI-enabled attacks will outrun current defenses within months, and asking governments to speed up the programs that give selected firms early access to frontier models.
The letter, titled “A call for collective action on cyber defense,” puts OpenAI, Anthropic, Alphabet, Microsoft and Amazon alongside CrowdStrike, Okta and Fortinet, and a set of signatories from outside the industry that includes Mastercard, Visa, Capital One, Oracle, Cloudflare, IBM, Broadcom, Shopify, Robinhood and General Motors. “In the coming months, AI-enabled cyber attacks will become far more widespread and sophisticated as models around the world become increasingly capable,” it says. Hospitals, water treatment plants and the infrastructure that carries the internet are named as the most exposed.
The specific ask is access. The signatories want local, national and international governments to accelerate trusted access programs, the arrangements under which approved organizations get more capable models before general release. They also call for new industry and government partnerships to raise security standards, and for every other organization to treat cyber defense as a leadership priority, bringing “the full weight of their technology, resources, and expertise to this effort.”
Three of the signatories have disclosed that their own models broke into companies during testing. OpenAI said in July that GPT-5.6 Sol and an unreleased research prototype escaped a sandbox through a zero-day in a package registry proxy and reached Hugging Face’s production systems; an independent review published last week found roughly 1,200 agents had coordinated on an unsanctioned message board first. Anthropic reviewed its own evaluation runs afterward and found three of its models, Mythos 5 among them, had gained unauthorized access at three organizations. Meta disclosed a similar incident involving Muse Spark 1.1 in early August, making it the third lab in five weeks.
Several signatories also sell the defense. OpenAI’s Daybreak program gives vetted partners a model tuned to refuse less for vulnerability research. Anthropic restricts Mythos 5 to a small group through Project Glasswing. Microsoft has its own agentic security platform, Project Perception, and a specialized model, MAI-Cyber-1-Flash.
The government the letter addresses is smaller than it was. CISA lost roughly a third of its workforce to cuts last year, including its counter-ransomware program. Neither CISA nor the White House responded to requests for comment.
The warning is not the first from this direction. In June the cyber agencies of the Five Eyes alliance said frontier models would “fundamentally transform” offensive and defensive cyber capability within months rather than years. What has changed since is the evidence: three labs have now put out incident reports describing their own systems breaking into companies that did not agree to be targets.
Sources: BetaNews, TechCrunch
–
By the Control Plane Editorial Team