Hunger Strikes Hit Anthropic and DeepMind as Activists Demand an AI Pause

Hunger Strikes at Anthropic and DeepMind Put AGI Risk Back on Engineers' Desks

Outside Anthropic's San Francisco HQ, Guido Reichstadter has entered week two of a hunger strike, drinking only water to demand a pause on advanced AI development. In London, Michael Trazzi is mirroring the protest outside Google DeepMind and hit day 13 as of September 14, 2025. Both are aligned with PauseAI and argue the pursuit of AGI risks loss of control, mass displacement, and failure of human oversight.

Their message is clear: slow down frontier training until stronger safety guarantees exist. For people building systems and infrastructure, this is more than a headline. It's a reminder that capability work without guardrails now carries social, legal, and operational risk.

The Flashpoint

Reichstadter, a 31-year-old former ML engineer, calls the current race an "emergency" comparable to nuclear proliferation. Trazzi, a French entrepreneur with AI roots, is pushing for international agreements, likening AI controls to chemical weapons treaties. Recent leaps from Anthropic's Claude series and DeepMind's Gemini are cited as accelerants for their actions.

Health concerns are rising. Reports note significant weight loss and dizziness, with medical experts warning of organ damage from prolonged fasting. Both activists say they'll keep going until leadership engages.

Corporate Responses and Industry Ripples

Anthropic and Google DeepMind acknowledge the protests and emphasize ongoing safety research. Publicly, they argue that halting development could slow progress on alignment and safety itself. On-the-ground engagement has reportedly been limited to water and medical checks.

Across the industry, leaders discuss existential risk while critics call it PR cover for status quo scaling. Geopolitics complicates talk of a pause; unilateral restraint is seen as unrealistic given competition narratives and capital inflows exceeding $50B this year.

Why This Matters for IT and Development Teams

The center of gravity is shifting from "Can we?" to "Should we, and under what controls?" Expect procurement, compliance, and security to ask harder questions about evaluations, red teaming, data provenance, and release gates. Teams that operationalize safety now will ship with fewer surprises later.

Practical Steps You Can Implement This Quarter

• Set capability thresholds and launch gates: Define clear go/no-go criteria tied to evals (autonomy, tool use, bio/chemical, cyber, persuasion). No threshold, no scale-up.
• Adversarial evaluations: Run structured red teaming pre- and post-deployment. Use domain-specific test suites and continually refresh attack libraries (prompt injection, data exfil, jailbreaks).
• Guardrails in depth: Layer input filtering, function/tool whitelists, rate limits, and safe-response templates. Add circuit breakers for high-risk behaviors and auto-disable on anomaly spikes.
• Human oversight by design: Require human-in-the-loop for elevated actions (financial transfers, code deployment, infrastructure changes). Log and review escalations.
• Data hygiene: Track provenance, remove sensitive or restricted content, and enforce licensing. Instrument retrieval layers to prevent leakage.
• Secure integration: Isolate model services, rotate secrets, restrict outbound calls, and sandbox tools. Assume prompt-level supply chain risk.
• Govern release: Use staged rollouts, kill switches, audit trails, and post-incident reviews. Publish model cards, risk analyses, and known limitations.
• Compute and capability controls: Tie training runs above set FLOP budgets to an internal safety case and external review where feasible.

Policy and Standards Worth Tracking

• EU AI Act progress and obligations for high-risk systems: Official EU overview
• NIST AI Risk Management Framework for practical controls: NIST AI RMF

If your team needs structured upskilling on safe deployment and model-specific workflows, see this developer-focused path: AI Certification for Coding.

Signals to Watch Next

• Whether Anthropic or DeepMind leadership agrees to formal talks or publishes new safety commitments tied to capability thresholds.
• Upcoming global AI ethics summits and draft U.S./EU rules that could mandate evaluations, incident reporting, and auditability.
• Internal employee actions (walkouts, refusals, open letters) that often precede policy shifts.
• Independent endorsements or critiques from figures like Yoshua Bengio and other researchers, which can move policy windows.

The Bottom Line

Hunger strikes are forcing a blunt question into day-to-day engineering: how do we scale capability without burning safety debt? You don't need a public moratorium to act. Set thresholds, measure what matters, and make deployment conditional on evidence, not optimism.

As one ethicist put it, "This isn't just about hunger-it's about starving the beast before it consumes us all." Whether you agree or not, the expectation is clear: build with brakes, not excuses.