Startups Build AI Verification Tech to Enable Global Slowdown

The trust problem in AI development

The United States and China face a strategic dilemma reminiscent of Cold War nuclear tensions: both want to avoid catastrophic AI risks, but neither can trust the other to honor a development slowdown. This impasse has kept frontier AI labs racing forward despite growing concerns about cybersecurity threats and autonomous weapons.

On June 12, the U.S. government took the unprecedented step of restricting access to Anthropic's Claude Mythos and Fable 5 models, treating them as cyberweapons too dangerous for foreign access. The move underscored how AI capabilities are increasingly viewed through a national security lens.

Major AI lab leaders have signaled openness to pausing development—if verification were possible. Anthropic stated in June that slowing AI progress "would likely be a good thing," but warned that an unverified pause "could leave everyone less safe" by letting reckless actors catch up. OpenAI CEO Sam Altman called for a new global organization capable of coordinating slowdowns when needed.

Why it matters

Without credible verification mechanisms, the AI race may continue accelerating toward potentially catastrophic capabilities regardless of what leaders say they want. The Cold War ended nuclear escalation through "trust but verify" monitoring—seismographs, satellites, and tamper-proof cameras that made treaty compliance observable. AI governance lacks equivalent tools, leaving the world stuck in an arms race dynamic even as the technology approaches capabilities that could destabilize cybersecurity and enable autonomous warfare.

Building cryptographic oversight

A small group of startups—roughly 50 people worldwide by one estimate—is developing verification technologies that could break the stalemate. These tools aim to enable oversight while protecting both industry secrets and user privacy.

Lucid Computing is building software that operates inside "trusted execution environments" on specialized AI chips from Intel and Nvidia. These secure enclaves can examine whether specific AI models are running or whether chips are training new models, then transmit only yes-or-no signals to outside observers. Nothing else about the computation would be revealed.

Kristian Rönn, Lucid's CEO, says this approach avoids the dystopian choice between global catastrophe and totalitarian surveillance. The company is testing its technology with U.S. government agencies and AI labs, though Rönn acknowledges it's not yet mature enough for international treaties.

British firm Amodo Design takes a different approach called recomputation, which re-runs portions of AI workloads to verify that data centers are operating agreed-upon models rather than secretly training more powerful ones. Co-founder Thomas Milton describes verification as "a ladder rather than a one-shot solution"—a stack of checks that grows more rigorous over time.

The harder problem: defining the target

Both technical approaches face limitations. Neither could detect a secret data center hidden underground. Training runs are "far easier to conceal than missile silos," Anthropic noted. Researchers at RAND argue that at least six different verification types may be necessary, including hardware security features, network monitoring, whistleblower programs, and traditional intelligence surveillance.

A deeper challenge looms: nobody has agreed on what exactly to verify. Unlike nuclear weapons, which require measurable uranium enrichment levels, AI capabilities resist simple quantification. Measurements are subjective and become outdated rapidly. Without clear negotiated restrictions, verification startups are aiming at uncertain targets.

Lennart Heim, an independent AI policy expert who coauthored the RAND paper, argues this makes AI governance primarily a policy problem rather than a technical one. "Nuclear weapons were easy by comparison," he says. "AI is not that. Not by any means."

China's willingness to participate remains uncertain, particularly given its position playing catch-up to U.S. labs. Beijing's 2025 criticism of Nvidia after U.S. lawmakers proposed remote chip-tracking capabilities suggests deep skepticism about Western-developed verification tools.

These details were first reported by AI Watch.