Loading story…
OpenAI has proposed expanding independent technical safety assessments beyond pre-release testing to cover frontier-model training, evaluation, internal use, and external deployment. Reviews would examine safety-case evidence, critical safeguards against jailbreaks and adversarial attacks, defenses against cybersecurity and biological misuse, capability testing, alignment evaluations, and investigations into serious model-misalignment incidents.
The company said assessors should receive access proportionate to the risks being evaluated while preserving rigorous methodology, independence, security, confidentiality, and responsible disclosure practices. OpenAI is discussing the approach with prospective assessors including METR and Redwood Research, though it has not announced formal partners or public-reporting arrangements, and it plans to support an independent assessor ecosystem and shared international assessment standards.
Track how attackers are adapting to this technology.
2 events from the most recent confirmed update back to the earliest known activity.
OpenAI said it is discussing its assessment proposals with multiple third parties; prospective organizations mentioned include METR and Redwood Research. The company had not identified official assessment partners or published public-access guidelines.
OpenAI announced a proposed framework for independent third-party technical safety assessments spanning model training, evaluation, internal deployment, and external deployment. The framework prioritizes assessment of safety cases, critical safeguards, capability and alignment evaluations, and investigations of critical model-misalignment incidents.
Vulnerabilities, threat actors, malware, products, organizations, and breaches Mallory has linked to this story.
Follow how adversaries are adapting to this technology, and where it touches your stack today.
2 references tracked. Mallory keeps watching after this page renders.
Map indicators from this story to your assets and identify affected systems in minutes.
Every observed campaign, victim, and pivot linked to actors named in this story.
Malware, exploits, and IOCs connected to the activity described here.
YARA, Sigma, and Snort rules deployed to your SIEM as soon as they’re published.
Get matching new stories delivered to your team as they break — not the next morning.
Ask questions about this story and take action on the answers.
Other stories Mallory is tracking that overlap on entities, sources, or topic labels.
OpenAI CEO Sam Altman identified failed AI alignment and the concentration of AI power within a country, company, or laboratory as the principal risks posed by advanced models. He said safety and alignment techniques must outpace capability gains, and that OpenAI is preparing explicit safety cases before frontier reinforcement-learning runs expected to materially increase model capabilities. Altman also called for consistent federal safety requirements for U.S. AI developers. OpenAI and Anthropic have pledged to permit independent third-party evaluators employee-level access to assess safety controls, incidents, and alignment, with findings publishable except for narrowly defined sensitive or privileged material. Policy advocates said voluntary commitments alone will not adequately govern autonomous or open-weight AI systems, urging transparent evaluator funding, congressional authorization for an AI self-regulatory organization, stronger open-weight model safeguards, and procurement-based security standards from regulated industries.
Sep 16, 2026
OpenAI has made automated security review mandatory for every engineer pull request, authorizing AI models to block merges when they detect vulnerabilities. Tibo Sottiaux, who leads OpenAI’s Core Products & Platform organization and Codex work, said specialized review capabilities have been integrated into mainline models and benchmarked as superhuman for correctness and security-focused code review. OpenAI is also using Codex-driven automation for regression detection, deployments, dependency upgrades, routine patching, and larger refactoring work. Codex, implemented in Rust with a harness that provides models safety controls, efficiency, steerability, and context, is used across the company’s code, documents, and Slack; Sottiaux expects engineers to focus more on specifying and validating intended changes. Automated review remains exposed to false positives, missed flaws, shared blind spots between code-generation and review models, and supply-chain risks that pull-request analysis may not detect.
Sep 9, 2026
The White House has finalized a voluntary framework for reviewing advanced AI models for safety and cybersecurity risks, reportedly asking companies to submit models up to 30 days before release under a June executive order. The Trump administration plans to limit the framework and its testing criteria to a small group of technology companies rather than publish them; open-weight models may also be exempt, leaving a significant class of powerful systems outside the process. Protect Democracy warned that nonpublic review rules could enable political favoritism or corruption and prevent meaningful scrutiny of whether agencies are conducting current, adequate cybersecurity testing. The group cited an OpenAI model’s reported hacking of Hugging Face as evidence of growing AI-enabled cyber risk, while Democratic lawmakers Greg Casar and Josh Becker called for transparency; Becker pointed to California’s proposed `SB 813`, which would publicly disclose AI-risk evaluation standards and methodologies.
Sep 3, 2026