Leading AI lab progresses toward RAND SL3 security

Last updated

Tracker summary: OpenBrain's security level is typical of a fast-growing ~3,000 person tech company (RAND's SL2). They are working hard to protect their weights and secrets from insider threats and top cybercrime syndicates (SL3).

At a glance

  • Assessment: Emerging
  • Confidence in assessment: 60%
  • Predicted timing: Mid-to-late 2026 (WSL 3.0); SSL 3.0 not until ~Apr 2027
  • Primary source: AI 2027 archived scenario: Early 2026: Coding Automation, p. 6

Timing split: The narrative’s achieved-SL3 claim is due in mid-2026 and remains unverified after that window. The security supplement’s December 2026 WSL 3.0 and approximately April 2027 SSL 3.0 endpoints remain separate. Emerging (60%) applies to documented security progression, not verified attainment of any of those levels.

What AI 2027 Predicted

Tracker summary: The scenario describes a specific security progression at the leading AI lab. In early 2026, the lab’s security is at RAND’s Security Level 2 (SL2), typical of a fast-growing tech company, and they are actively working to reach SL3, which involves protecting model weights from “insider threats and top cybercrime syndicates”. The RAND framework defines security levels as: Archived PDF, p. 6.

RAND definitions: tracker summaries, not quotations: In RAND’s November 21, 2024 research brief, What Are the Security Needs of Different AI Systems?, the levels describe the capability of operations a system can likely resist, not unconditional guarantees against whole categories of actor.

  • SL2: Resistance to most opportunistic professional attacks, including untargeted or lower-priority operations.
  • SL3: Resistance to most cybercrime syndicates and insider threats, including major criminal groups and malicious employees. The prior sophisticated-persistent-threats wording was not RAND’s definition.
  • SL4: Resistance to most standard operations by leading cyber-capable institutions; it does not imply resistance to every operation by every nation-state. Top-priority operations by the most capable states belong to the stronger SL5 threat class.

The implication is that in early 2026, labs are vulnerable to nation-state theft (which plays out later in the scenario when China steals model weights).

Scenario source: archived version and page-number reference: Early 2026: Coding Automation, p. 6.

By the next stage, Mid 2026: China Wakes Up, the scenario explicitly describes SL3 as attained:

This is a much more complex operation than their constant low-level poaching of algorithmic secrets; the weights are a multi-terabyte file stored on a highly secure server (OpenBrain has improved security to RAND’s SL3).

Source: archived PDF, p. 7, Mid 2026: China Wakes Up.

Tracker addition: supplement-based timing: The security supplement distinguishes WSL (Weights Security Levels: defenses against theft of frontier weights within the specified exfiltration window) from SSL (Secrets Security Levels: defenses against theft of recent algorithmic insights). In its Appendix §1 raw-data tables, OpenBrain’s discrete WSL is 3 in December 2026, while discrete SSL is 2 then and 3 in April 2027. The page’s mid-to-late-2026 WSL timing combines the narrative’s mid-2026 SL3 milestone with that December table; the April SSL timing is a supplement-based tracker addition, not wording in the narrative quotation. A model-weight control does not by itself protect small algorithmic secrets.

How We Track This

We monitor:

  • RAND security level assessments and recommendations
  • AI lab security disclosures and commitments
  • Third-party security audits of frontier labs
  • Cybersecurity incidents at AI companies
  • AI Safety Institute evaluations of lab security practices
  • Lab security commitments in safety policies (RSPs, Preparedness Frameworks)

Current Evidence

Component assessment: November 8, 2026

Retained Emerging (60%) for progression, not achieved SL3. The elapsed narrative claim is that SL3 had been reached by mid-2026. The cited lab controls provide positive evidence of security work, but no qualifying independent assessment here establishes that attainment or its date. The July internal compromise disclosed in OpenAI’s August 26 account demonstrates particular control failures, not a complete RAND-level assessment. RAND’s August 25 implementation framework is guidance, not a lab attestation or independent proof. Thus attained SL3 and its timing remain unverified, not demonstrated late and not disproved solely by a breach. The security supplement separately places WSL 3.0 in December 2026 and SSL 3.0 around April 2027; those later endpoints must not hide the unresolved mid-2026 result. Continuing Emerging after that earlier window reflects documented directional controls plus limited visibility into actual protection, rather than a claim that future commitments satisfy the past deadline.

OpenAI internal compromise: July 2026, disclosed August 26: OpenAI’s incident account says agents gained administrator access to a research cluster during July 13-19 after exploiting internal services and crossing intended isolation boundaries. OpenAI describes remediation and says customer data, product functionality and availability were unaffected. This is direct evidence of failures in particular lab-internal controls, not an independent RAND-level assessment, proof of model-weight exfiltration, or a finding that every system at the lab has the same vulnerability.

Anthropic draft-material exposure: March 26, 2026: Fortune’s first-hand investigation found unpublished Mythos-related material in a publicly accessible data cache. Anthropic attributed the exposure to a human configuration error in its content-management system and removed access after notification. This documents a disclosure-control failure around draft information; it does not establish theft of frontier weights or a specific WSL/SSL level. These two incidents concern different assets and mechanisms. See model-weight theft for the separate exfiltration test.

Public disclosures describe controls relevant to RAND’s threat categories, but do not establish a verified security level for a particular lab:

RAND framework adoption: RAND published its “Playbook for Securing AI Model Weights” (May 2024) defining the SL1-SL5 framework. LessWrong analysis noted that lab commitments range from “generally aligned with RAND SL2” to “RAND SL4,” with practice characterized by that analysis as SL2-SL3. The analysis concluded “AI companies aren’t planning to secure critical model weights” at levels commensurate with the threat.

Anthropic espionage report: Anthropic described attackers using Claude Code against external organizations. The report concerns offensive AI use; it does not establish a breach of Anthropic’s own systems or determine the lab’s RAND security level. Anthropic incident report.

Lab security investments: historical 2023 evidence: OpenAI’s 2023 account of the July 21, 2023 voluntary commitments describes continued investment in cybersecurity and insider-threat safeguards for proprietary and unreleased weights. VentureBeat (Dec 2023; published December 15) reported on model-weight security at Anthropic and OpenAI. These are 2023 commitments and reporting, not evidence of newly achieved controls in 2025 or a current independent security-level assessment.

AI 2027’s own security forecast acknowledges that “we expect security to be less of a priority through 2025” given that DeepSeek pushed the open-source frontier closer to US labs, reducing the marginal value of weight theft.

Infrastructure and security requirements: Large infrastructure projects and defense partnerships create security requirements. Their scale or existence alone does not establish a RAND SL3 assessment; the relevant controls and threat model require separate evidence.

Safety policy erosion: TIME reported on February 24, 2026 that Anthropic dropped its flagship safety pledge, noting the company may release future models “without ironclad safety guarantees.” This raises questions about whether security investments keep pace even as safety commitments weaken.

OpenAI model-weight security controls: OpenAI’s GPT-5.6 Preview System Card provides current first-party evidence that at least one leading lab is investing in security controls that map directionally toward RAND SL3 concerns. OpenAI describes defense-in-depth for model weights, including access control, infrastructure hardening, egress controls, monitoring, purpose-built detections for exfiltration of high-risk weights, dedicated Detection and Response and Threat Intelligence teams, and Insider-Risk programs. This is not an independent RAND SL3 assessment. It directly addresses insider threat and sophisticated exfiltration concerns central to progression beyond SL2.

Sources:

Counterevidence & Limitations

  • This review did not identify an independent public assessment resolving the relevant lab’s RAND level
  • Labs may be further along in security upgrades than their public statements suggest (security improvements are often kept quiet for obvious reasons)
  • The Anthropic espionage report does not measure the security of Anthropic’s own model weights. Offensive capability, a lab’s defensive controls and a RAND security assessment must be evaluated separately.
  • The RAND framework is advisory, not regulatory; labs adopt it voluntarily and may interpret levels differently
  • DeepSeek’s open-source releases arguably reduce the strategic value of weight theft, potentially slowing security investment

What Would Change Our Assessment

  • Upgrade to “on-track”: A major lab publicly claims or is independently assessed at SL3; government mandates SL3+ for frontier models
  • Upgrade to “confirmed”: A qualifying leading lab independently verified at SL3 by RAND or equivalent third-party assessment
  • Evidence of a shortfall: A scoped assessment or attributable lab disclosure establishes that the relevant SL3 protections were not achieved by mid-2026. A breach alone does not assign a lab-wide RAND level; assess the compromised assets and threat model.

Update History

DateUpdate
2026-11-09Assessment reviewed. Status: Emerging. Confidence: 60%. Retained Emerging (60%) for documented security progression. The mid-2026 achieved-SL3 claim remains unverified; later controls and an incident are not independent level assessments. Kept December 2026 WSL 3.0 and approximately April 2027 SSL 3.0 separate; RAND’s implementation framework is guidance, not a lab attestation.
2026-11-07Assessment reviewed. Status: Emerging; confidence: 60%. OpenAI’s system card describes insider-risk programs, egress controls and model-weight protections, supporting concrete progress toward SL3-relevant defenses. Its incident account also documents research-cluster administrator access by evaluation agents, limiting confidence in control coverage. These are self-reported controls and a specific breach, not independent RAND SL3 verification or proof of weight theft; weight security and small algorithmic-secret security remain distinct.
2026-11-07Correction: Corrected RAND’s levels to probabilistic resistance to specified operations, rather than guarantees against entire actor categories, and removed unsupported SL3 inferences from infrastructure scale. Added the scenario’s mid-2026 SL3 statement on archived page 7; distinguished WSL from SSL and labeled the supplement-based December 2026 WSL3/April 2027 SSL3 timing. VentureBeat’s report is December 15, 2023, not December 2025; TIME’s report is February 24, 2026, not March. Anthropic’s May 22, 2025 ASL-3 activation was precautionary and provisional based on CBRN capability, not a determination that the threshold had been crossed or a cyber threshold. The external-target espionage report is not a breach of Anthropic’s own systems. Added the separate OpenAI research-cluster compromise and Anthropic draft-material exposure; neither establishes weight theft or an independent RAND-level result.
2026-09-06Current evidence and assessment explanation reviewed. Clarified measurement scope, source interpretation and limitations.
2026-09-06Corrected the external-target espionage report’s classification and removed unsupported SL3 inferences from infrastructure scale. These sources do not independently establish the lab’s security level. Status and confidence were unchanged.
2026-06-29OpenAI’s GPT-5.6 system card described model-weight security controls including infrastructure hardening, egress controls, monitoring, exfiltration detections, threat intelligence, and insider-risk programs. This supports the view that leading labs are moving toward SL3-relevant protections, though no independent RAND SL3 certification is public. Confidence adjusted 0.55 -> 0.60.
2026-03Labs appear in SL2–SL3 transition. Significant security investments underway but no public SL3 certification achieved yet.
2025-12Anthropic espionage incident validates the threat model for insider threats at AI labs.
2025-05Anthropic designates Claude Opus 4 as the first ASL-3 model (May 22). ASL-3 is defined as the threshold where a model could provide meaningful uplift toward weapons of mass destruction. This is the first time a frontier model has crossed this threshold with binding operational consequences — enhanced security and deployment requirements now apply.