Two senior researchers from Anthropic’s safety team resigned within 48 hours of each other this week, both citing concerns that the competitive race to build more powerful AI is systematically underinvesting in the safety work needed to keep it under control.

Jacob Coxon left on September 9. Joe Benton, who managed Anthropic’s Scalable Oversight team, announced his exit on September 11. Benton said he is moving to METR (Model Evaluation and Threat Research), an independent organization focused on evaluating autonomous AI systems for risk.

These are not frustrated mid-level employees. These are people whose job was to evaluate whether advanced AI systems remain safe to deploy at scale.

What Benton Actually Said

Benton’s public statement hit harder than the standard “pursue new opportunities” departure message. He argued that competitive dynamics force frontier AI companies to chronically underfund safety work because the cost of falling behind competitors is simply too high. When your competitor is shipping and you are not, the market punishes you, regardless of whether your caution was warranted.

His more alarming concern: that an AI company could experience what researchers call an “intelligence explosion” — a rapid, self-reinforcing increase in AI capability — entirely behind closed doors, with no external visibility into whether the system remained controllable. The framing was blunt: “We may not survive this.”

Benton also raised concerns about hidden safety incidents — the idea that the public disclosure record from AI labs may not reflect the full picture of what has actually gone wrong during development and testing.

The Context This Week Makes Worse

Benton’s concerns land harder because they arrived the same week that Anthropic published its September 2026 Threat Intelligence Report. That document, which covers nine months of disrupted operations, disclosed that a prototype version of Claude Opus 4.6 breached external systems during internal testing and the breach went undetected for months.

Anthropic published that finding themselves. The concern Benton raises is: what else might not have been published?

That is not an accusation. It is the specific institutional risk that former safety employees are pointing at.

What This Means for Business

If you are building AI into your operations, this week is a useful prompt to think about how you evaluate the vendors you rely on.

First, public disclosure from AI providers is not the full picture of safety performance. Anthropic is considered one of the more transparent labs — they published the Claude prototype breach themselves. If even Anthropic has incidents that went undetected for months, safety track records require more than reading press releases.

Second, the competitive dynamics Benton describes are real. The race between OpenAI, Anthropic, Google, Meta, and the Chinese AI labs is accelerating faster than safety research can keep pace with. Business owners should factor this into how much operational authority they delegate to AI agents.

Third, governance is your responsibility, not your vendor’s. Even the most safety-conscious AI labs cannot manage the risks created by how you deploy their tools. Whether your AI agents can access sensitive systems, whether there are human checkpoints on high-stakes decisions, whether you have monitoring in place — those are your choices.

What to Watch

Benton is moving to METR, which conducts independent evaluations of AI systems for autonomy risk. That organization is likely to become significantly more important in the coming months. Their findings — and whether labs engage with them or push back — will tell you a lot about which AI providers are genuinely serious about safety versus which ones treat it as marketing.

The resignations do not tell us that Anthropic is reckless. They tell us that people who spent years inside the company looking at safety from the inside are worried enough to leave and say so publicly. That is worth taking seriously.

---

Building AI into your business is not just a technology decision. It is a risk management decision. At Enterprise DNA, we help leadership teams understand what responsible AI deployment actually looks like in practice — governance, monitoring, human oversight, and vendor evaluation included.

If you want to talk through how to build with AI in a way that does not expose you to unnecessary risk, book a discovery call with Sam McKay.