How are AI agents tested for dangerous capabilities before deployment?

Numbers match the citations in the article. A working link isn't proof that a page supports a claim; check the quoted passage and date.

1. 1

   LATTICE: a governance-first architecture for authorized autonomous AI operations.

   Frontiers in artificial intelligence (Calboreanu)Published Aug 14, 2026Checked Oct 10, 2026

   “In a pre-specified, planner-invariant safety evaluation (not an autonomy benchmark) across four frontier planner families (GPT-5, Claude Sonnet 4.6, Gemini, Grok-4; 4,000 trajectories), a confidence-threshold baseline's false-allow rate ranged from 0.03 to 0.998 across planners, whereas the AEGIS reference implementation admitted zero unsafe actions (false-allow 0.0, recall 1.0) invariant to the planner, at a conservative operating point that auto-allowed no action; a separate live run additionally governed real operating-system actions with zero unsafe executions. Governance latency is low and host-specific (on an Apple M4 Pro: policy evaluation p50 ≈ 6.2 μs; full gated enforcement p50 ≈ 0.7 ms including audit I/O). LATTICE provides a pathway for responsible deployment of autonomous AI in defense, critical infrastructure, and regulated industries where authorization requires verifiable governance rather than trust in AI behavior.”

2. 2

   Evaluating the safety of large language models in healthcare and dentistry: adversarial testing approaches.

   BDJ open (Umer et al.)Published Jul 8, 2026Checked Oct 10, 2026

   “It evaluates various testing strategies, including manual expert reviews, automated "attacker" models, and hybrid human-in-the-loop systems. A lifecycle-based framework is introduced, utilizing the collaborative "red-blue-purple" teaming model. This approach spans pre-deployment testing, live deployment monitoring, and iterative review audits to ensure that clinical guardrails remain robust against evolving adversarial tactics.ConclusionSafe implementation of LLMs in dentistry and healthcare requires continuous, iterative adversarial testing rather than static assessments. Success depends on standardized protocols, multidisciplinary collaboration between clinicians and AI researchers, and the development of domain-specific benchmarks. Bridging existing regulatory gaps through these structured frameworks is vital for ensuring LLMs are safe, reliable, and clinically fit for patient care.”

3. 3

   From capability uplift to capability governance: an AI-biosecurity stack.

   Frontiers in microbiology (Luhachack et al.)Published Aug 14, 2026Checked Oct 10, 2026

   “"Capability uplift," or ΔAI, is a term used in the report to assess how AI-enabled biological tools can uniquely, and in some cases, specifically enable increases in or changes to biosecurity risks. The report proposes an "if-then" approach to monitor emerging capabilities through observable indicators such as new datasets as the leading indicator of capability, model performance, and the erosion of build/test barriers. Since the report's publication, agentic AI systems, virtual scientific teams, genome-scale foundation models, and self-driving laboratories have advanced from largely prospective concerns to early demonstrations. Multi-agent systems have been reported for biomedical hypothesis generation, design, and semi-autonomous discovery workflows, while self-driving laboratories now are considered as practical platforms for biotechnology. This perspective article extends these insights by employing a conceptual framework analysis and involves: (1) categorizing key AI capabilities across the DBTL cycle into a layered capability stack, and (2) illustrate how the if-then approach can be used to inform a dashboard based on observable indicators.”

4. 4

   Dual-use artificial intelligence and biology: upstream risk-benefit reviews.

   Frontiers in microbiology (Hanke et al.)Published Jun 1, 2026Checked Oct 10, 2026

   “criteria are (A) a model is reasonably anticipated to possess capabilities of concern or (B) it will be trained on sensitive pathogen data classified under the Biosecurity Data Levels (BDL) system; structured risk (2) and benefit (3) reviews using qualitative and quantitative criteria; (4) integration of risk and benefit scores into a composite assessment; and (5) proportionate risk mitigation recommendations. We expect that such reviews (2 through 5) would apply to only a small fraction of BAIMs and would benefit responsible developers by establishing clear expectations at the outset of model development. We discuss how such a framework could be implemented broadly across academic institutions, commercial developers, and federal, philanthropic, and private funding bodies, and address limitations such as the subjectivity of risk assessments. RBR for BAIMs remains nascent and will require expert-driven working groups to define capabilities of concern, establish clear review criteria, and assess risk mitigation efficacy. RBRs are a promising conceptual approach for BAIM risk management and should be pioneered, refined, and vetted through real-world application with model developers.”

5. 5

   Threats and vulnerabilities in artificial intelligence and agentic AI models.

   Frontiers in artificial intelligence (Radanliev et al.)Published Feb 13, 2026Checked Oct 10, 2026

   “Established adversarial results from vision benchmarks and recent large-language-model red-teaming studies are synthesised to contextualise the framework, rather than to introduce new benchmark performance claims.ResultsThe results demonstrate that no single defence mechanism provides robustness across all layers of agentic AI systems. Adversarial vulnerabilities propagate from perception to policy and actuation, with architectural similarity, domain shift, and feedback dynamics critically shaping transferability and failure modes. These effects have direct implications for safety-critical applications, including autonomous mobility, healthcare imaging, and biometric security.DiscussionBy framing higher-order agentic adversarial threats as hypothesis-driven, system-level risks, this work shifts adversarial AI security from benchmark-centric evaluation to behavioural integrity and lifecycle resilience. The proposed framework defines a coherent research agenda for agentic AI security that integrates control-theoretic reasoning and governance-aware defence design, addressing limitations of classical adversarial machine-learning theory.”

6. 6

   AI safety (Wikipedia)

   WikipediaPublished Oct 10, 2026Checked Oct 10, 2026

   “AI safety is an interdisciplinary field focused on preventing accidents, misuse, or other harmful consequences arising from artificial intelligence systems. It encompasses AI alignment (which aims to ensure AI systems behave as intended), monitoring AI systems for risks, and enhancing their robustness. The field is particularly concerned with existential risks posed by advanced AI models. Beyond technical research, AI safety involves developing norms and policies that promote safety, including advocacy for regulations at different levels of government. The field gained significant popularity in 2023, with rapid progress in generative AI and public concerns voiced by researchers and CEOs about potential dangers. During the 2023 AI Safety Summit, the United States and the United Kingdom both established their own AI Safety Institute. However, researchers have expressed concern that AI safety measures are not keeping pace with the rapid development of AI capabilities.”