On September 9, Jacob Coxon posted on X that he had resigned from Anthropic, accusing both labs of “racing straight to self-improving superintelligence and gambling with our lives” — a post that reached 70 million views. Within hours, an Anthropic alignment lead publicly posted his personal risk estimate and stated Anthropic lacks a superintelligence alignment plan. Two days earlier, OpenAI’s chief scientist published a personal essay calling for voluntary industry slowdowns.

For operators, this is not a doomsday signal. It is a new variable in vendor planning.

Key takeaways:

• Coxon worked in pretraining research at both labs; his departure signals internal tension over safety pacing.
• Jakub Pachocki, OpenAI’s chief scientist, published a personal essay — not a formal company policy — calling for voluntary slowdowns to “become commonplace.”
• Pachocki wrote that OpenAI will “unilaterally withhold further scaling as needed” — his personal position, not a company notification commitment.
• Evan Hubinger, still at Anthropic, personally stated a “>10% within the next decade” risk estimate and that Anthropic lacks a superintelligence alignment plan — his views, not Anthropic’s position.
• Operator posture: add one safety-gate question to your next AI vendor review.

What Actually Happened at Both Labs?

Coxon worked on pretraining at both companies. The post gained traction when Evan Hubinger, still at Anthropic, replied with two personal claims: a “>10% within the next decade” risk estimate, and that Anthropic “does not yet have a plan to solve alignment for superintelligence.” These are Hubinger’s own views, not Anthropic’s official position. Neither company has formally responded to Coxon’s accusations.

Pachocki’s September 7 openai.com essay is personal writing, not a formal OpenAI policy statement. He argued no company has solved alignment sufficiently “to continue responsibly scaling at maximum speed for much longer,” called for voluntary slowdowns, and wrote that OpenAI will unilaterally withhold scaling when needed.

Why Does This Change Your AI Vendor Assumptions?

Three signals, three planning implications.

Coxon’s departure is a talent signal — not a predictor of delays. Patterns matter more than individual events.

Hubinger’s post matters because it comes from inside Anthropic — a sitting alignment lead stating the alignment plan is incomplete is not external criticism.

Pachocki’s essay adds the most direct planning variable. OpenAI has already disclosed a training pause and pacing criteria; Pachocki’s essay frames this as potentially recurring — including unilateral scaling withholding. No formal enterprise notification policy has been announced.

What Operators Should Do Now

The near-term move is not to exit either vendor. Both are shipping current products under existing contracts.

Add one question to your next vendor review: If a safety evaluation triggers a release pause, how will you notify enterprise customers — and does your agreement include any commitment on that?

Watch for: formal release-gate criteria from either lab; customer-notification language in enterprise agreements; OpenAI disclosing a training resumption date; or a second named departure from either lab’s safety team.

For safety-gating context, see our briefings on GPT-6 Astra’s enterprise rollout and the Daybreak access program.

---

Frequently Asked Questions

Should operators stop using Anthropic or OpenAI because of these departures?

No. A researcher resignation and a personal essay do not change contracts, product availability, or service reliability. The signal is forward-looking — future releases carry a safety-gate variable acknowledged but not yet formalized into enterprise notification policy. Add sharper vendor questions, not platform decisions based on a social media event.

What does “voluntary slowdown” mean for my AI product roadmap?

Pachocki’s call appears in a personal essay, not a formal policy announcement. OpenAI has already disclosed a training pause. If slowdowns formalize, expect delays to next-generation models, not disruption to current ones. Build 2027 roadmaps with a buffer for possible safety-gate delays.

How should safety researcher departures factor into vendor due diligence?

Treat them as a monitoring signal. Signals worth escalating to procurement review: a pattern of named departures from safety or alignment teams, formal changes to published evaluation timelines, or safety-gate language in enterprise agreements.

---

Sources: Jacob Coxon X post, September 9, 2026; Evan Hubinger X post, September 9, 2026; Jakub Pachocki, “An Alien Mind,” OpenAI, September 7, 2026; CNBC expert roundup, September 9, 2026.