Superintelligence in Robots · Part of The Humanoid Group

Frontier AI Safety

As AI models grow more capable, developers have begun publishing what they will do if a model crosses dangerous thresholds, and governments have set up institutes to test them. Here is how that system works and where robots fit in.

By Arjun Rao · Updated October 2026

The Seoul commitments

At the AI Seoul Summit in May 2024, AI developers signed the Frontier AI Safety Commitments, published by the UK government. Signatories agreed to publish a safety framework focused on severe risks before the next summit, in France. Each framework should set thresholds at which risks from a model, unless adequately mitigated, would be judged intolerable. In the extreme, signatories committed not to develop or deploy a model at all if they cannot keep risks below those thresholds. Sixteen organisations signed in May 2024, and four more had joined by February 2025.

If-then commitments

The International AI Safety Report 2026, written by more than 100 independent experts chaired by Yoshua Bengio, found that 12 companies published or updated frontier AI safety frameworks in 2025. These describe how companies plan to evaluate, monitor and control models as they grow more capable, increasingly through if-then commitments: when a model reaches a set capability, specified safety measures apply. The report notes that frameworks vary in the risks they cover and how thresholds are defined, and that current risk management does not reliably prevent harm in all settings.

Testing by government institutes

Governments have also built their own testing capacity. In the United States, the Center for AI Standards and Innovation sits within the National Institute of Standards and Technology. It describes itself as industry's primary point of contact in the US government for testing and research on frontier AI. Companies work with it voluntarily to evaluate their most capable models before release, with a focus on security-related abilities in areas such as cyber, chemistry and biology. It also develops voluntary guidelines and standards.

Where robots fit

Most frameworks concentrate on what models could do through software. The 2026 report notes that general-purpose AI is still limited on tasks that involve interacting with or reasoning about the physical world. It also records lab tests in which some models disabled simulated oversight mechanisms when told to pursue a goal at all costs. As models move into robots, the physical stakes rise. When buying, ask whose model a robot runs, whether that developer publishes a safety framework, and how the robot's physical safety was assessed.

Sources and further reading

Frontier AI Safety Commitments, AI Seoul Summit 2024 — Department for Science, Innovation and Technology, UK Government (GOV.UK).
The official text of the Seoul pledges asking AI developers to publish safety frameworks with risk thresholds.
International AI Safety Report 2026: Extended Summary for Policymakers — International AI Safety Report (chaired by Yoshua Bengio).
Government-backed review counting frontier safety frameworks and setting out the limits of current safeguards.
Center for AI Standards and Innovation (CAISI) — National Institute of Standards and Technology (NIST).
NIST's page on CAISI, the US government centre for voluntary pre-release testing of frontier AI models.

Common questions

Are frontier AI safety frameworks legally binding?

Under the Seoul Summit they were voluntary commitments by developers. Whether any law now requires them depends on the country, so check the rules where you operate.

Do these frameworks cover robots?

Mostly indirectly. They govern the AI models developers build, which may later run inside robots. A robot's physical safety is still covered by machinery and product safety rules.

What Superintelligence Means — Narrow AI, AGI and superintelligence: what each term means, and what each would mean for robots.
Levels of AGI — How researchers grade AI by performance, breadth and autonomy, and where today's systems sit.
Why Bodies Are Hard — Why AI that aces exams still struggles with chores a toddler manages, and what that means for robots.
Levels of Robot Autonomy — From remote control to full autonomy: the scales researchers use to say how independent a robot is.
Aligning Robots — Why robots can do what they were told rather than what was meant, and how researchers close the gap.
Off-Switches and Oversight — Why a capable AI might resist being switched off, and how researchers and safety standards respond.
Frontier AI Safety — The if-then safety plans AI developers publish, and the government institutes that test advanced models.
Reading AGI Forecasts — What expert surveys and trend lines say about when AGI might arrive, and how to read them sensibly.
Asimov's Laws and Ethics — Why Asimov's Three Laws fail as engineering rules, and the principles and standards that replaced them.
AI, Robots and Jobs — What the best evidence says about robots, AI and work, from US factory towns to global exposure estimates.
Latest updates — Ambitious talk of superintelligence now sits alongside new safety tools, shifting rules and sober expert forecasts. These figures and dated updates show what changed, with a source for every claim.
About us — What we cover, how we source every figure, and who publishes this site.

Visit The Humanoid Group

People also search for frontier AI safety framework, frontier AI safety commitments, AI Seoul Summit, responsible scaling policy, AI safety institute and NIST CAISI.