Guidelight AI Standards is an independent nonprofit AI-safety standards organization co-founded by Page Hedley and Steven Adler, both formerly of OpenAI’s safety staff; Adler serves as Chief Scientist (Source: clear-eyed.ai; guidelight.ai). It publishes assessable safety standards for frontier AI companies; its first, the Control standard v1.0, was released in May 2026 (Source: clear-eyed.ai), and its catalog also lists Capability Testing v1.0 and Transparency v1.0 (Source: guidelight.ai).
Control standard v1.0
On July 15, 2026, Guidelight published an explanation of the Control standard v1.0. The standard sets six minimum principles for preventing loss of control during internal deployment of advanced AI systems:
- Tamper-evident activity logs
- Misbehavior scanning
- Stress-testing of defenses
- Proactive barriers
- Independent third-party verification
- Breach-response planning
Guidelight said in July 2026 that it was conducting its first assessment of leading AI companies against the standard (Source: clear-eyed.ai); that assessment was published the following month and is described below. The standard text itself was published May 19, 2026 (Guidelight AI Standards — Control (v1.0, May 2026)). The standard codifies, as third-party assessable requirements, the deployment-protocol approach developed in the AI control research agenda.
Control assessment of frontier AI companies (August 2026)
Guidelight published its first assessment of frontier AI companies against the Control standard in August 2026, with information current through August 18. It assessed Anthropic, Google, Meta, OpenAI and xAI, and stated its overall finding as: "basic practices for keeping control of AI are, at most, partially implemented" (Source: guidelight.ai).
The assessment drew only on public materials — system cards, safety frameworks, risk reports, blog posts, and third parties' descriptions of collaborations with companies — and narrowed the Control standard to six practices, which it named differently from the standard's own six principles listed above: logging of internal AI activity; measurement of monitor efficacy; gated actions, meaning high-risk actions a monitor must clear before they take effect; circuit breaking, meaning a temporary halt after a surge of flagged misbehaviour; third-party review; and a containment plan for a misaligned model. Each company received a 0–5 absolute score per practice and a relative rating against peers; the overall grade is the average of the six absolute scores (Source: guidelight.ai).
| Practice | Anthropic | OpenAI | xAI | Meta | |
|---|---|---|---|---|---|
| Overall grade | C+ (2.50) | C+ (2.50) | D+ (1.50) | D− (0.83) | F (0.67) |
| Logging | 3 | 3 | 2 | 0 | 1 |
| Monitor efficacy | 3 | 3 | 1 | 0 | 1 |
| Gated actions | 3 | 2 | 1 | 2 | 0 |
| Circuit breaking | 3 | 2 | 1 | 2 | 0 |
| Third-party review | 3 | 2 | 2 | 0 | 2 |
| Containment plan | 0 | 3 | 2 | 1 | 0 |
Scale: 0 not implemented · 1 precursors only · 2 limited partial implementation · 3 substantial partial implementation · 4 near-full implementation · 5 full implementation. Information current through August 18, 2026 (Source: guidelight.ai).
No company scored above 3 on any of the six practices, and Guidelight reported that a majority of scores were 2 or lower. It grouped the five into three tiers: Anthropic and OpenAI with the strongest practices, both logging substantial portions of internal AI activity and scanning it for misbehaviour while fully implementing none of the assessed practices; Google lagging in implementation but holding the most specific published forward-looking document, its AI Control Roadmap, which Guidelight said it found little public evidence of Google having implemented; and Meta and xAI weakest, with Meta's controls known primarily through its disclosures to METR for the May 2026 Frontier Risk Report and xAI the only assessed company that did not participate in that report (Source: guidelight.ai).
Guidelight identified detection and third-party assessment as the strongest areas across the five companies, and prevention and containment as the weakest. On prevention — gated actions and circuit breaking — only Anthropic scored above "limited partial implementation," which Guidelight said leaves control systems liable to being disabled by misbehaving AI or overwhelmed by a faster sequence of attacks than the company can answer. It also stated that companies' future plans for third-party assessment are unclear on frequency and on depth of access, and that only Anthropic appears to have granted hands-on access to assessors to date (Source: guidelight.ai). That question of evaluator access is the subject of We Must Pace the Frontier (Amodei, September 2026), published the following month, which proposes employee-level access for third-party evaluators.
The single largest divergence within the table is containment planning, where Anthropic scored 0 and OpenAI 3 — reversing their otherwise matched positions and leaving the two companies tied on the overall grade for different reasons (Source: guidelight.ai).
Organization
Page Hedley is co-founder and chief executive and Steven Adler co-founder and chief scientist; Janine Edgerton-Avin is chief operating officer and Erez Coopfer chief financial officer (Source: guidelight.ai).
Relationships
- instance-of: AI Control (codification as an assessable standard)
- related: Guidelight Control Standard (v1.0) (the standard’s own page)
- related: Steven Adler (founder), Frontier AI Governance, Future of Life Institute (parallel third-party company-assessment efforts such as the AI Safety Index)
- related: METR — the Frontier Risk Report of May 2026 is a principal evidence base for the August 2026 assessment's scoring of Meta, and non-participation in it is recorded against xAI
- supports: We Must Pace the Frontier (Amodei, September 2026) — the assessment's finding that only one of five companies has granted hands-on access to third-party assessors is evidence for that essay's case for embedded evaluators, though the two are independent
Confidence note: Medium — the assessment methodology is now published in full alongside the scores, and the assessment has been reported on by third parties. Funding remains unreported, and the organization's own publications (guidelight.ai, clear-eyed.ai) are still the source for most of what is known about it.