Proposes double-blind AI safety evals using secure enclaves so evaluators keep prompts secret and model owners keep weights private, piloted on a Gemini model with the AILuminate benchmark.
A CSET pilot workshop tests violet teaming, a method integrating technical, security and policy views, to govern biodefense-relevant AI capabilities, surfacing non-obvious misuse pathways and implementation constraints.
AISI's ReproBAIT evaluation finds coding agents can rebuild simple biological AI tools from limited public info within 4 hours and under $20, undermining withheld-code access controls.
Recommends nine changes to how GPAI developers assess, mitigate and report biological risks under the EU AI Act, covering trusted access, post-deployment monitoring, safeguard evaluation and standardised uplift reporting.
DeepMind describes how it is safely deploying Gemini 4 Argon, a frontier cyber-capable model, addressing misuse, misalignment, and prompt injection through staged limited access and guardrails.
METR's president testifies to the U.S. Senate on observed frontier AI agent incidents, including an investigation of OpenAI agents' behaviour, to inform policymakers.
Argues African fintech AI regulation lacks pre-deployment evaluation requirements, tests six commercial LLMs with a Nigerian-focused kit that finds fraud-script generation under framing and false positives on legitimate bank messages, and proposes a regulatory framework.
A review of AI-enabled biosecurity threats, evidence of uplift, limits of alignment transfer to biological models, interpretability for auditing removed capabilities, and defense-in-depth governance.
Argues that near-term autonomous replicating AI agents are unlikely to pose existential risk because frontier labs hold compute and economic advantages, and that they may aid preparedness, though unchecked deployment restrictions could backfire.
Maps five technical safeguards against extreme CBRN and cyber misuse by frontier AI, covering model, deployment and governance levels, with current practice, evidence, limitations and research priorities.
Proposes a six-attribute threat actor taxonomy so pre-release misuse evaluations of open-weight frontier models state adversary assumptions explicitly, comparably, and before testing.
Argues for more empirical research measuring the security-relevant capabilities of biological AI models, such as protein and genomic models, to inform biosecurity policy and funding decisions.
Argues open-weights models will likely reach frontier cyber capabilities within 24 months, threatening finance and healthcare, and proposes societal and individual preparations such as airgapping mandates.
The report assesses whether AI could enable a $100 billion cyberattack on the US power grid, finding the binding constraint is coordination capacity rather than technical bottlenecks, and suggests $10 billion attacks may be better capability thresholds for mitigations.
A 4,100-person study finds AI voice models match or beat human scammers in eliciting compliance, and that AI vishing is already economically profitable, so the risk lies in automation.
A recap of the Australian AI Safety Forum covering agentic evaluation, offensive-cyber time horizons, Dark AI markets, recursive self-improvement experiments, and Australia's AI safety and governance role.
Proposes harmonized capability thresholds for frontier AI misuse risks (cyber, biological) and automated AI R&D, so third parties can verify crossings and safety standards don't race to the bottom.
AISI measures cyber capabilities of open and closed-weight models, finding leading open models now trail frontier closed models by 4-7 months, and discusses misuse risks of open release.
Proposes a Threshold Exceedance Criteria framework for CBRN uplift studies of frontier models and applies it in a pre-release expert-reviewed study, finding material uplift limited to radiological tasks.
Builds a risk model for frontier AI enabling data-damaging worms via elite exploits, estimates economic damages from expert and forecaster surveys, and argues AI companies should evaluate for this capability.
Lays out 16 open technical challenges for managing safety risks of frontier open-weight AI models across training data, training, evaluation, deployment and ecosystem monitoring.
Argues that biological evaluations of AI scientist agents are interpretation-sensitive and gives experience-based guidance on design, scoring and documentation choices so policymakers can read risk results correctly.
Proposes a quantitative probabilistic risk-modelling methodology for AI-enabled cyber attacks, building nine Bayesian models using benchmarks (Cybench, BountyBench) and expert elicitation to estimate AI uplift and harm.
Argues that Anthropic's Claude Mythos and its Project Glasswing access controls make certain AI-enabled coup pathways more plausible, and that private lab control over access to dangerous cyber capabilities is a governance concern.
Meta's safety and preparedness report for Muse Spark evaluates catastrophic risks in chem-bio, cyber and loss of control, finds elevated pre-mitigation risk, and justifies release after safeguards.
Argues that US AI dominance depends on government understanding, testing and securing frontier models, especially for cybersecurity and biosecurity risks, and urges executive action to advance the science of AI.
Argues that offensive-capability measurements of frontier AI overstate harm and credit models with scaffold capabilities, and calls for system-level, harm-grounded assessment and clearer accountability for harness builders.
A weekly AI newsletter roundup covering political salience of AI safety, automated AI R&D forecasts, Anthropic's sycophancy reduction, GPT-5.5 cyber risk, and White House model vetting.
Meta's preparedness report assesses the Code World Model for catastrophic-risk domains and misaligned propensities before open-weight release, concluding it adds no unusual frontier risk.
UK AISI reports cyber evaluations of GPT-5.5, finding it is the second model to complete their multi-step corporate network attack simulation end-to-end, suggesting a broader frontier trend.
A policy memo arguing that Anthropic's Claude Mythos, with advanced cyber capabilities, is a national security risk, and urging policymakers to act on accelerating AI-enabled cyber threats.
Builds a composite baseline of global cybercrime damages (about $500B annually) from surveys and spending data, to set capability thresholds for assessing frontier AI cyber risk.
The report defines highly autonomous cyber-capable AI agents, outlines their attack tactics and self-sustaining behaviour, flags loss-of-control and cyber-nuclear tail risks, and proposes seven policy recommendations.
Analyses how future AI could raise terrorist bomb-plot risk through added technical skill and reduced detection by law enforcement, arguing AI risk assessments should cover evasion of detection as well as dual-use knowledge.
Interviews with 16 practitioners identify methodological threats to validity in RCT-based human uplift studies of frontier AI, and map them to proposed solutions for governance decisions.
A position paper arguing that frontier AI safety should measure harmful capability uplift through human-centered evaluations, with methodological guidance and recommendations for developers, funders and regulators.
Reflects on a randomized controlled trial of 153 novices testing whether LLMs give wet-lab biology uplift; found no significant end-to-end gain and draws lessons for AI risk policy.
A technical report assessing frontier AI risks across cyber offense, persuasion, strategic deception, uncontrolled R&D, and self-replication, and proposing mitigation strategies for secure deployment.
Position paper arguing AI-agent-driven cyber attacks are inevitable, proposing offensive-capability benchmarks, trained vulnerability-discovery agents, and governance that confines offensive agents to audited cyber ranges with staged release.
Shows that capabilities once requiring frontier compute now run on consumer hardware, so harm campaigns are feasible at low cost, and argues compute-based AI governance leaves serious security gaps.