- The Five Eyes intelligence alliance issued a joint warning that frontier AI models are approaching “cyber-lethal” capability against critical infrastructure, with officials telling governments the threat timeline is “months, not years,” according to Axios.
- Jan Leike, who co-led OpenAI’s Superalignment team, resigned in May 2024 citing safety de-prioritisation and joined Anthropic to build a new alignment science team, a move the research community widely read as a public statement about where serious safety work could be done.
- Ilya Sutskever’s Safe Superintelligence raised over $1 billion in its first funding round despite having no product or revenue model, a credibility signal that institutional investors are treating the safety-versus-capability tension as a material concern, not a philosophical one.
The Five Eyes intelligence alliance has warned that the most advanced AI models are approaching “cyber-lethal” capability against critical infrastructure, with a threat timeline of “months, not years,” according to Axios. The warning landed against a backdrop of high-profile departures from major AI labs by researchers who say safety concerns drove them out, including some who built the very models the alliance is now flagging.
Five Eyes Sounds Alarm on Frontier AI Threats
According to Axios, the warning identified advanced models from Anthropic as a leading cyber-lethal threat, with OpenAI‘s models identified close behind. Officials also flagged the accelerating pace of AI development in China and Japan, where cheaper models have advanced faster than agencies initially anticipated.
The framing is unusually direct for an allied intelligence assessment, Axios noted. Characterising a commercial AI model as “cyber-lethal” in a shared bulletin represents a notable escalation in public posture from the alliance. Whether the assessment leads to coordinated regulatory action or remains a warning on paper is not yet clear.
Talent Exodus Over Safety Prioritisation
For well over a year before the Five Eyes bulletin, researchers inside the major frontier labs had been raising versions of the same alarm internally, and, when that failed, publicly. Jan Leike, who co-led OpenAI’s Superalignment team, resigned in May 2024, stating he had disagreed with OpenAI leadership about the company’s core priorities and that safety had “taken a backseat to shiny products.” He subsequently joined Anthropic to build a new alignment science team.
Leike’s departure was not a quiet exit. By making his reasoning public, he put internal disagreements about safety prioritisation on the record in a way that individual resignations rarely do. His move to Anthropic was read across the research community as a statement about where serious alignment work could actually be done.
Co-Founders Launch Safety-First Superintelligence
Ilya Sutskever, OpenAI’s former chief scientist and co-founder, left the company in May 2024 after nearly a decade. The following month, he announced Safe Superintelligence (SSI), co-founded with former OpenAI engineer Daniel Levy and former Apple AI lead Daniel Gross. SSI raised over $1 billion in its first funding round. The company’s stated mandate is narrow by design: develop safe superintelligence, with a business model structured to keep safety and security insulated from commercial product pressures.
The size of that first raise is worth noting. Over $1 billion for a company with no near-term product, no revenue model and a single declared goal is a credibility signal about how seriously institutional investors are taking the safety-versus-capability tension. Whether SSI can maintain that insulation as investor expectations accumulate is a different question, one the company has not yet had to answer.
Anthropic’s Foundation on Safety Concerns
The pattern has earlier precedents. Anthropic itself was founded by researchers who left OpenAI over safety and governance concerns. CEO Dario Amodei, along with several colleagues including researcher Tom Brown, departed OpenAI in late 2020 and early 2021, according to reports. Their stated concern was the direction OpenAI was taking as it shifted to a capped-profit model and deepened its partnership with Microsoft. Anthropic was built, at least in its founding narrative, around a stronger commitment to AI safety and interpretability research.
That context complicates the Five Eyes warning. One of the models the alliance reportedly flagged as a cyber-lethal threat was built by a company whose founding argument was that existing labs were not taking safety seriously enough. The gap between safety-first founding principles and the capabilities those same organisations eventually produce is a recurring tension in this space, and the intelligence assessment puts it in sharp relief.
The Stakes of the AI Safety Race
Leike stated publicly that safety had taken a backseat to products at OpenAI. Sutskever structured SSI explicitly to insulate safety work from commercial pressure. The Five Eyes bulletin, according to Axios, reflects an intelligence community conclusion that the risks those researchers flagged internally have reached a threshold requiring a public allied warning. For context on the broader legislative response to these risks, see our coverage of how Congress is codifying human responsibility for AI decisions. For more coverage of AI policy and regulation, visit our AI Policy & Regulation section.
