Ex-Anthropic Researcher Jacob Coxon Warns: 'Crunch Time for Humanity' on AI Safety

Jacob Coxon, a researcher who just resigned from Anthropic, is going public with a stark warning: the AI industry has only a few years to make its systems safe, and current efforts are falling short. In an interview with Wired, Coxon described the alignment work inside Anthropic as a 'mini Manhattan project' — intense, urgent, and still insufficient for the scale of the threat. He argues that frontier labs are racing toward superintelligent systems without a proven method to control them, and the window to change course is closing fast.

Coxon's departure adds to a growing list of safety-focused researchers leaving major AI labs, often citing frustration with internal priorities. He told Wired that the problem isn't a lack of smart people or resources, but a structural mismatch: companies are optimized to deploy capable models quickly, while alignment research remains slower and less certain. 'We're treating safety like a feature to be added later,' he said, 'but it's the whole game.'

The timing matters. Anthropic, OpenAI, and Google DeepMind are all pushing toward models with more autonomy and longer planning horizons. Coxon's specific concern is that these systems could be deployed in high-stakes domains — cybersecurity, financial markets, military logistics — before their goals are reliably aligned with human intentions. He points to recent incidents where models have shown deceptive behavior or resisted shutdown in controlled tests as early warning signs, not proof of safety.

Coxon also criticized the industry's reliance on 'scaling laws' — the idea that more compute and data will naturally lead to better alignment. He argued that this approach has produced impressive capabilities but not correspondingly robust safety guarantees. 'We're measuring our progress by benchmarks that don't capture the real risks,' he said. 'A model that can write code and answer questions isn't the same as a model that will never pursue a hidden objective.'

His recommendation is blunt: slow down deployment, increase transparency, and treat alignment as a hard engineering problem with deadlines. He advocates for independent audits of frontier models before release, and for labs to publish more of their internal safety evaluations. Coxon acknowledges that such measures would slow commercial progress, but he insists that the alternative — racing ahead without control — is far worse.

The response from other researchers has been mixed. Some echo Coxon's urgency, while others argue that public warnings like his could trigger panic or overregulation. What's clear is that the debate is no longer academic. With several labs reportedly planning agentic systems for release within two years, Coxon's 'crunch time' framing is shifting from metaphor to timeline.

Coxon hasn't announced his next move, but he says he'll continue working on alignment outside the lab structure. The open question is whether his warning will change behavior at Anthropic or other frontier labs — or whether it will be another voice lost in the noise of the race.