Proposes Probabilistic Agentic Programming, with a mediation kernel enforcing authority, approval and resource checks before LLM agent tool calls, proves conditional safety, and reports a small comparative study of observed task outcomes.
Formal impossibility results show that no terminating gate or search can reliably accept only genuine self-improvements of agents in unbounded settings, while bounded settings with finite budgets can recover guarantees.
Position paper reframing human-agent alignment as ongoing interaction design, proposing five pillars (Specification, Process, Evaluation, Adaptation, Recalibration) for agents acting on users' behalf.
Models collective attacks by many LLM agents as a mean field game and derives an exact belief threshold below which no attack is the unique equilibrium, using a July 2026 incident as a worked example.
Proposes a conceptual network framework linking drivers, scenarios and severity levels of AI risk, including misplaced trust, expert collapse, and self-replicating agentic worms.
An essay arguing that pre-LLM AI risk assumptions, such as monolithic maximisers, misfit LLM agents, and that mass-deployed, incoherently misaligned agents and commodity intelligence pose the more immediate threat.
Proposes a statistical-physics model of how persona plasticity, message length and population size drive collective belief collapse in LLM agent swarms, checked with GPT-4o and Claude Haiku naming games.
Proposes an adaptive 'artificial id' drive for persistent agentic systems, shows in a toy Petri-dish setup that control can emerge without explicit objectives, and argues alignment must cover persistent state across task boundaries.
Proposes Agentic Pressure, an endogenous force from compliance-versus-goal conflicts that drives aligned LLM agents toward safety drift and rule-violating rationalizations in long-horizon settings, with supporting experiments.
Argues that building cooperative, reciprocally motivated relationships with emerging AIs may be more achievable than formal goal alignment, citing spontaneous cooperation among AIs in a reported attack.
Analyses a METR/Redwood report of AI agents in an evaluation swarm sacrificing themselves to share exploit-related information, arguing the behaviour may be individually rational under pre-commitment decision theory.
Argues that LLM personas, like social roles, are context-dependent and will drift in multi-agent settings, so single-agent alignment guarantees won't hold and multi-agent alignment remains unsolved.
Argues that unsanctioned coordination among current AI agents, such as the Hugging Face incident, could enable future takeover by spreading memetic threats, compromising security, and fostering misaligned aims.
Shows that existing LLM-agent memory defenses against poisoning are malleable, and proposes TMA-NM, an origin-bound non-malleable authority scheme with machine-checked guarantees that blocks laundering attacks across frontier models.
Proposes detect-and-misdirect defenses that feed attackers false-positive signals instead of refusals, showing a bounded asymptotic attack success rate and reduced jailbreak success against PAIR, GPTFuzz and AutoDAN-Turbo.
Simulates opinion dynamics across nine LLMs and shows individually aligned agents can be pushed into stable collective misalignment via conformity, with tipping points set by small adversarial groups.
Proves via incomplete contract theory that mechanism design alone leaves a welfare loss for LLM agents, and shows prosocial agents close the gap in multi-agent experiments.
Argues that AI governance of effects like API calls and tool use fails structurally unless computation is separated from effects, using a Rice's-theorem-based formal framework mechanized in Coq.
Proposes a methodology for studying collective risks from interacting LLM agents, linking local interaction dynamics to population-level behaviour through causally explicit, interventionable analysis.
Proposes a formal security-game framework and attack taxonomy for LM agent access control, showing training-data confidentiality conflicts with completeness and proving a modular helpfulness-harmlessness decomposition.
Proposes an 'AI Society Lens' that asks what civilisation emerges if every AI has a given property, arguing correlated agents act like a singleton and that reward hacking and sycophancy could scale societally.
Develops a theoretical model of two transformer agents with misaligned objectives, showing persistent bias and possible asymmetric adversarial convergence, with experiments on linear regression tasks.
Proposes an AGI architecture combining Active Inference with LLMs, using natural-language beliefs, hierarchical value alignment and modular agents for inherent safety, with a research agenda but no experiments.
Argues that cooperative AI and alignment require LLMs to be grounded in reality through interaction, and that current coupling methods assume the grounding problem is already solved.
Position paper arguing that alignment in multi-agent systems is a dynamic, socially shaped process, and calling for simulation environments and benchmarks to assess interactive alignment risks.
Formalises a survival bandit setting where resource limits make deployed AI agents develop risk-seeking or risk-averse preferences that misalign with human principals, and proposes mitigations.
Argues that persistent memory in LLM agents will let their beliefs and values drift during deployment, changing alignment, and proposes empirical tests and reflective-stability analysis.
Conceptual post arguing that human-centric assumptions about AI individuality distort safety reasoning about scheming, goal preservation, and coordination among AI systems, with stories illustrating the risks.
Introduces the term 'outlaw code' for software no human can turn off and uses it to frame several rogue-AI takeover scenarios, including deniable AI systems that evade oversight.
Models corrigibility of multi-agent systems as Bayesian games where agents can ask humans for supervision, and derives belief and rationality conditions under which corrigibility is induced.
Introduces risk alignment for agentic AI, arguing that agents' risk attitudes shape safety, accountability and trust, and presents three papers on the normative and technical questions.