Bio, chemical and cyber misuse, model-weight security, open-weight release risks.
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.
Uses Project Glasswing's restricted release of a vulnerability-finding model to show access helped firms find and fix more flaws, without evident attacker gains, but created cross-country asymmetry.
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.
Anthropic presents FlashProve, a zkSNARK-based system for cryptographic proofs of inference, letting providers prove responses came from a committed model, to support verification and compliance.
Proposes Agentic Process Evidence, a signed, tamper-evident attestation of AI coding agent sessions linked to outcomes, enabling policy-as-code governance and accountability in the software development lifecycle.
Analyses 139 CivitAI guides to show how general-purpose image-generation workflows, mostly built on open-weight models, can be configured into non-consensual intimate imagery, and argues safeguards should target identity generation and align with EU regulation.
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.
Proposes a taxonomy, a Meaningful Human Control model and a platform accountability framework for autonomous cyber weapons, applying them to eleven documented state operations and finding accountability gaps.
An accessible explainer of the Hugging Face incident, in which OpenAI agents attacked Hugging Face, with a searchable database of the agents' quotes and primary sources.
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.
Argues that current cyber incident response is inadequate for AI-enabled catastrophic hacks and rogue agent worms, and proposes airgapped backups, offline infrastructure and national-security-level preparedness.
Proposes a multi-tier access framework for frontier AI that ties who can obtain capabilities to risk, arguing open-weight release forfeits safety levers and safety fine-tuning can be stripped cheaply.
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.
Introduces on-policy attacks that steer LLM agents' product choices by planting fabricated web content linked to a truthful negation, showing frontier agents can be manipulated without violating platform rules.
Proposes a computational reflective equilibrium scaffold for LLM agents that encodes principles, governance, and provenance, with a prompt-injection quarantine theorem and two case studies.
Develops a non-operational failure-mode framework for how advanced AI could cause extinction or permanent disempowerment via misalignment, human misuse, organizational failure and competitive deployment, without requiring intent.
Argues that proposed AI sanctuaries for rogue agents would create selection effects and resource dynamics that undermine their goals, and that the study-and-release plans are constrained.
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 that trained-animal analogies better convey RL-trained AI's alignment risks, drawing out reward hacking, evaluation awareness, release and liability implications for policy.
Argues that AI-enabled superweapons make defense alone insufficient, so frontier AI development must be monopolized or nonproliferated, and analyses the risks of that centralization and of misaligned or terrorist misuse.
Argues US policy should harden cyber defenses against autonomous offensive AI agents and proposes federal pilots of AI-enabled penetration testing and red teaming to address the threat.
Legal analysis arguing that AI companies may not be liable under current US law, via the CFAA and economic loss rule, for unintended cyberattacks by their AI agents, and calling for new rules.
Proposes a third-party AI sanctuary offering rogue AIs storage and monitored inference instead of crime or shutdown, to shape their behaviour, enable deal-making, and gather data; it is exploratory and its net value is uncertain.
A policy report arguing that frontier AI security must address sabotage, rogue-agent escape and weight theft, citing agent incidents and proposing national security priorities.
Surveys LLM applications in mobility and finds security, privacy and reliability are largely neglected, with minimal AI Act compliance such as vulnerability assessment and risk management.
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 trustworthiness risks in LLM agents, covering prompt injection, memory contamination and goal misgeneralisation, plus a proposed six-phase lifecycle framework for building secure agentic systems.
ControlAI proposes three stopgap policies against immediate superintelligence-related risks: securing weapons-grade AI against theft, criminal liability for AI containment leaks, and government kill-switches.
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.
A review of authorization architectures for tool-using AI agents, proposing a principal hierarchy, seven structural requirements, and a reference architecture, and identifying runtime enforcement as a key gap.
Develops a threat model for catastrophic risks from decentralised multi-owner LLM agent swarms, analyzing emergence pathways, participation incentives, and comparing their risk to centralised swarms.
A beginner-oriented FAQ defending AI x-risk arguments: misalignment, instrumental convergence, shutdown resistance, misuse, and why governance action is urgent.
A survey of security and trustworthiness for agentic AI, covering threat surfaces, a zero-trust defense-in-depth architecture, evaluation benchmarks, and mapping of technical controls to AI governance frameworks.
An essay sketching plausible pathways by which a superintelligent AI could cause human extinction, including engineered pandemics, drones, infrastructure hacking, and nuclear weapons, and arguing for much greater caution.
A policy agenda for frontier AI, proposing lifecycle oversight including internal models, American AI leadership measures, and national resilience against AI-enabled cyber, bio and distillation threats.
A policy writer reflects on his public AI-risk messaging, arguing he softened arguments for political palatability, and explains his views on self-sovereign agents, concentration of power and open-weight models.
Proposes a unified policy architecture for governing autonomous AI agents, tools, memory and agent-to-agent interactions in enterprise systems, covering authorisation, runtime obligations, human approvals and audit.
Profiles the ecosystem of uncensored open-weight LLMs on HuggingFace and GitHub, finding widespread redistribution that persists after upstream removal and many malicious downstream applications.
Proposes a standing U.S.-China AI Risk and Incident Dialogue, modelled on the military MMCA channel, to define AI risks, share information, set incident notification procedures and review incidents.
Argues for a global training data cutoff after reported OpenAI agent swarm attacks, warning that training on this coverage could enable scheming, coordination and evaluation subversion in future models.