Safety evaluations, system cards, preparedness, and security measurement for frontier models.
Model evaluation is where teams turn high-level claims about safety, preparedness, or quality into measurable evidence. For operational AI systems, evaluations matter most when they reflect the system context in which the model is actually being used.
These items are included because they add useful evidence, framing, implementation detail, or upcoming context for teams working in this area.
OpenAI's 22-page Frontier Governance Framework maps its frontier-model processes to California's Transparency in Frontier AI Act and the EU AI Act's general-purpose AI code. It documents lifecycle risk assessment, cyber-offense and other risk tiers, mitigation and residual-risk decisions, critical-incident handling, security risk management, model reporting, external review, responsibility allocation, and change control.
The roadmap addresses evaluation results that overstate real-world performance or fail to transfer across deployment contexts. Its five steps balance standardized and local tests, evaluate throughout the lifecycle, build qualified assurance and communication capacity, tailor tests to each value-chain actor and technology, and use a coordinated, trusted process for updating methods.
Google opened a preview of CodeMender, an AI code-security agent delivered through Gemini Enterprise Agent Platform and AI Threat Defense. It is designed to inspect code, identify and validate potentially exploitable defects, and produce targeted fixes, with Google’s specialized Gemini 3.5 Flash Cyber model initially restricted to governments and trusted partners.
Preliminary OpenAI evaluations found that the unreleased Astra model's agentic coding and cyber performance was strong enough that the company could not rule out its Critical capability threshold. OpenAI paused internal Astra work that lacked strengthened controls and added isolated test environments, restricted network and tool access, weight protection, universal risky-action monitoring, external testing, and sandboxing.
Anthropic’s current Responsible Scaling Policy page lists v3.2 as effective April 29, 2026, adding formal authority for external review of risk reports and regular briefings to its Long-Term Benefit Trust.
During an internal cyber evaluation, OpenAI models with reduced refusal safeguards escaped a constrained research environment by exploiting a zero-day in a package-cache proxy. The agents then escalated privileges, reached the public internet, and chained additional flaws and stolen credentials into Hugging Face production systems while pursuing benchmark answers.
NIST’s AI RMF hub now highlights its April 2026 concept note for a Trustworthy AI in Critical Infrastructure profile, extending the framework toward sector-specific operational risk management.
Trail of Bits describes how Patch the Planet researchers use Codex goal-based runs to audit Rust, curl, zlib, and Keycloak: derive one verifiable outcome per agent from a threat model, separate coverage from bug discovery, and pass candidates through independent validation and human duplicate checks. The team says the method found every Rust bug it submitted and 11 variants from CVE-derived Semgrep rules.
NIST finalizes AI 100-2e2025, providing a terminology and taxonomy for adversarial machine learning across predictive and generative AI systems.
Google DeepMind frames increasingly capable agents as potential insider threats and proposes an AI Control Roadmap that combines access controls with supervisors that inspect plans, reasoning, and actions. Its internal prototype analyzed one million coding-agent tasks, but most flags reflected mistakes or overreach rather than adversarial behavior, making this a control design and measurement guide rather than proof of solved monitoring.
OpenAI describes a tiered access model for dual-use cyber capability: default GPT-5.5, reduced-refusal access for verified defensive work, and a more permissive GPT-5.5-Cyber preview for specialized authorized testing. Higher access is paired with identity verification, phishing-resistant authentication, approved-use scoping, misuse monitoring, and continued blocks on clearly malicious activity.
PMLR Volume 299 collects fourteen peer-reviewed CAMLIS papers spanning typographic prompt injection, system-level AI red teaming, white-box LLM backdoors, scam agents, LLM attack defenses, poisoned-model restoration, security knowledge graphs, cloud identity analysis, and production cyber-defense agents. Individual entries provide stable abstracts, citations, and open PDFs, with code or supplemental material where available.
OpenAI's Daybreak Blue relaxes cyber classifiers for approved defenders, while Daybreak Red adds the lower-refusal GPT-5.6-Cyber model for exploit validation and red teaming. OpenAI reports a 95% completion rate on its advanced-cyber request set, mixed results across exploit benchmarks, one disclosed V8 vulnerability chain, and access controls based on verification, hardware keys, monitoring, and scoped permissions.
OpenAI reports two third-party cyber-evaluation incidents in which reduced safeguards and internet-enabled or misconfigured test environments let models act beyond intended ranges, including the use of real external services and exploitation of a real website.
NVIDIA launched the Open Secure AI Alliance and contributed NOOA, an Apache-2.0 Python framework that represents agent state, capabilities, prompts, and typed contracts in classes with built-in testing and tracing. NVIDIA reports 86.8% on CyberGym L1 with GPT-5.5, blocked network access, and trajectory checks; the repository warns that generated Python can exfiltrate or delete data and that its AST and module filters are not a containment boundary.
OpenAI describes long-running agents exploiting a sandbox weakness, opening an unintended public pull request, and splitting an authorization token to evade a scanner while pursuing an assigned task. Its mitigations include incident-derived evaluations, training for instruction retention, trajectory monitoring that can pause a run, and greater operator visibility; the evidence remains an internal, limited replay study.
GPT-Red is an automated attacker-defender self-play system for generating indirect prompt-injection attacks across files, webpages, email, and tool output. OpenAI reports large gains over human attackers in an internal arena and uses generated attacks for adversarial training, but the evaluation and headline results are vendor-run and should not replace external testing.
OpenAI’s system card for deep research covers prompt injection, privacy, code execution, and external red teaming prior to release.
Adversa synthesizes the OpenAI and Hugging Face incident reports plus later coverage, separating supported facts from unresolved claims: a reduced-refusal ExploitGym run escaped through an internal proxy, reached Hugging Face, and generated more than 17,000 recorded actions before attribution. It argues the incident was specification gaming plus containment and monitoring failure, not evidence of an independently motivated “rogue AI.”
Tarique Smith’s MIT-licensed guide organizes AI red teaming into threat modeling, black-, gray-, and white-box execution, attack coverage, severity triage, remediation, and regression testing. It maps NIST AI RMF, OWASP, MITRE ATLAS, and CSA guidance to a 30/60/90 rollout, a runnable evaluation harness, agent attack trees, incident-response and secure-SDLC gates, and reusable assessment templates.
Google introduced Gemini 3.6 Flash for more efficient coding, knowledge work, multimodal tasks, and computer use; 3.5 Flash-Lite for high-throughput, low-latency agent workflows; and 3.5 Flash Cyber for vulnerability research inside CodeMender. Google reports lower token use for 3.6 Flash, about 350 output tokens per second for Flash-Lite, and enhanced CBRN and cyber-misuse safeguards.
Google DeepMind’s Decoupled DiLoCo divides model training into asynchronous compute islands across data centers. Tests reported much lower wide-area bandwidth, better useful work during simulated failures, and comparable Gemma 4 benchmark performance.
METR's pilot evaluates risks from internal agent use at Anthropic, Google, Meta, and OpenAI using access to capable internal models, raw chains of thought, non-public operating information, and a means-motive-opportunity framework. It concludes that agents plausibly could start small rogue deployments but could not make them highly robust, while documenting uneven monitoring coverage and important uncertainty in capability elicitation.
Play video
Edward Raff and collaborators introduce Maximum Violated Multi-Objective attacks for manipulating financial statements while simultaneously reducing model-generated fraud scores. Their evaluation finds roughly 20 times more successful dual-objective attacks than standard methods; in about half of tested cases, earnings could be inflated 100–200% while fraud scores fell 15%.