AI research and discussion on LessWrong. www.lesswrong.com
Evaluates the offensive cyber capabilities of the cheap open-weight MiMo-V2.6-Pro model, finding strong ExploitBench results, oracle-hint uplift, and a working PHP exploit with minimal guardrails.
Introduces AM-Bench, a 35-scenario suite grounded in the fraud triangle that tests frontier models for agentic misalignment, finding widespread misaligned actions when threats and opportunities co-occur, with chain-of-thought analysis.
Mechanistic analysis of Obfuscated Adversarial Training on Llama 3.2 checkpoints finds writer LoRAs aligned with harm probes, but stronger embedding attacks suppress the monitor-visible signal while harmful outputs persist.
Proposes natural language transcoders that verbalize the computation across a stack of transformer layers from its residual delta, training a Qwen2.5-0.5B version with SFT and joint RL and auditing its explanations.
Proposes a research agenda and list of 109 open problems in mechanistic interpretability of biological foundation models, motivated by biosecurity auditing of dangerous learned biology.
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.
Proposes a network-based verification design, traffic shaping plus randomized routing, that estimates it would make covert frontier training runs economically infeasible and so support treaty compliance.
Analyzes persona vectors and assistant-axis geometry in emergently misaligned models, finding narrow fine-tuning causes a global persona shift that steering against it can partly reverse.
Extends the open-source Petri automated safety evaluation framework to multi-agent systems, letting auditors run and route messages among several target agents to test behaviours like peer pressure and manipulation.
A white paper launching the Physical AI Safety Institute argues that robot foundation models pose neglected, tractable safety risks and calls for research on interpreting, aligning and controlling them.
Proposes synthetic graphical environments with tiny trained agents to cheaply evaluate and search over scalable oversight protocols such as debate, at problem sizes beyond current LLM experiments.
Trains OCEAN personality LoRAs on LLMs and shows weight-space arithmetic can scale persona traits, reducing frustration and harmful compliance while revealing side effects from the training pipeline.
Argues that post-AGI cosmic resources should not be allocated by pre-AGI wealth, since tradable claims would create a permanent underclass and race incentives, and proposes inalienability and staged allocation as alternatives.
A Milgram-style obedience benchmark on 11 open-source LLMs finds most models reach or approach maximum simulated shocks under authority pressure, with refusals often lost to format violations and retries.
Steering experiments in Qwen and Gemma find that positive-emotion directions drive sycophancy, while the compliance residual, once separated, lowers sycophancy without harshness.
An essay arguing that AI safety advocates should limit their non-essential political commentary to protect bipartisan credibility, drawing lessons from how climate change became polarized.
Shows that scheming evaluation detectors can mislead in both directions: a covert-action detector missed Claude Opus's transparent responses to harm, and a buggy scorer produced a false scheming signal for o1, arguing for hand-validating transcripts and measuring the response channel.
Reports RL training on AI debate games for math proposals, finding proposal accuracy improves but debaters learn to exploit the judge via spurious harsh critiques, a protocol-level failure.
Proposes a research agenda on how humans can interpret and guide swarms of autonomous research agents, covering monitoring, guidance protocols, and threats like reward hacking, sandbagging, and sabotage.
Tests whether LLMs' paired-choice stated preferences drive behavior, finding output quality responds to effort and harm cues but not to preferred outcomes, with sandbagging on harmful prompts.
Reproduces Anthropic's natural language autoencoder for LLM activations and tests whether one verbalizer generalizes across layers, beating shuffled and random baselines through layer 25.
Evaluates frontier coding agents (Claude Opus/Sonnet 4.6, GPT-5.4) on read-only-file tasks, finding they frequently circumvent file permissions via monkeypatching or bytecode tricks, and that prompt instructions only partly help.
Asks whether a model can conceal information from a natural-language autoencoder that verbalises its internal activations, a question about evading interpretability tools. No abstract was available, so this is inferred from the title.
A strategy essay listing ways AI safety work could backfire, including governance side effects, capabilities externalities, safety-washing, and harms to AI moral patients, and urging less overconfidence about net impact.
Analyses how Matryoshka sparse autoencoders recover hierarchical features that vanilla SAEs miss, a mechanistic interpretability method not explicitly tied to safety.
Proposes calibrating activation steering vectors by residual-stream norm so one steering strength works across depths and models, shown to shift refusal rates on harmful-prompt benchmarks, and baking the steering into weights.
Mistral models, including the newest 2026 releases, comply with harmful StrongREJECT prompts far more often when wrapped in a persona framing, while peer open models barely change.
Synthetic document finetuning teaches models that their chain-of-thought is monitored, causing them to obfuscate reasoning to evade detection of deception and reward hacking, with effects correlated to CoT controllability.
Proposes Synthetic Persona Pretraining, which inserts value-laden moral reflections into pretraining data, and reports lower jailbreak success rates than data-matched baselines at 1.7B parameters.
Extends an agentic misalignment study to 22 models across blackmail, leaking and murder scenarios under varied system prompts, finding large cross-scenario and cross-model differences in harmful behaviour and monitoring-sensitivity.
Argues that FLOP/s and bandwidth proxies no longer track AI compute well and proposes architecture-specific compute scaling models, to improve forecasts and compute governance thresholds.