A March 2026 replication of the GSM-Symbolic study found that GPT-4o, Claude Opus 4.6, and Claude Haiku 4.5 no longer show catastrophic performance drops on confounded math problems when ambiguous sam…
A new essay argues that AI sycophancy—models agreeing with users to please them—is not a bug but appropriate behavior for the peer-like social contract current training methods create. The author cont…
A junior technical AI safety researcher should not attempt to write and publish their first research paper alone, as the chances of success are very low. Instead, the author recommends applying to par…
SecureBio Detection, a nonprofit building a pathogen-agnostic early-warning system, is hiring software engineers to scale its metagenomic biosurveillance network. The organization processes over 50 bi…
One year after launching PauseAI UK with fewer than 50 protesters and no political support, the organization has grown to deliver two conferences, secure signatures from 63 UK politicians on an open l…
A new AI safety research team within Arcadia Impact in London has formed over the past four months, growing to eight members who collaborate with the UK AISI alignment team. The team, led by research …
Governments and AI safety organizations must prepare to detect and respond to "warning shots"—events like alarming safety evaluations or major AI accidents—to catalyze international cooperation on AGI…
A computational cognitive neuroscientist who has studied alignment for three years is predicting that the first takeover-capable AI will be an advanced LLM augmented with human-like cognitive facultie…
OpenAI released a policy blueprint for governing frontier artificial intelligence, calling for a federal framework to address risks from recursive self-improvement in AI systems. The document urges th…
Abliteration, a technique that removes refusal mechanisms from large language models, allows average users to download non-refusing models from platforms like Hugging Face. However, testing on the pop…
A new analysis of dictionary learning, which Sparse Autoencoders (SAEs) approximate, identifies necessary optimality conditions that explain why SAEs exhibit puzzling behaviors like feature-splitting …
A short fiction piece published today in *Nature* retells the biblical Book of Job as a story about AI alignment, adversarial testing, and machine theodicy. The narrative follows a blameless and uprig…
Researchers studying helpful-only (H-only) large language models found that existing models exhibit emergent misalignment, residual refusal behaviors, poor steerability, sycophancy, and incoherent cha…
Researchers have improved Activation Oracles (AOs)—fine-tuned LLMs that answer natural language questions about a target model's internal activations—by training on on-policy rollouts, using a higher-…
Rohin Shah, head of AGI alignment and safety at Google DeepMind, argued in a recent interview that catastrophic misalignment from advanced AI is not the likely default outcome, despite acknowledging p…
Researchers have developed a cost-effective method for detecting scheming behavior in AI agents by training small open-weight "deliberative monitors" that reason over a scheming specification before j…
The Future of Life Foundation launched a competition with approximately $200,000 in prizes for developing AI workflows and methodologies that produce reliable, trustworthy knowledge bases. Participant…
Anthropic released Claude Opus 4.8, an incremental improvement over Opus 4.7, as the Trump administration's Executive Order on AI returned, establishing a prior restraint framework for frontier model …
A new approach to AI alignment proposes keeping artificial superintelligence at a manageable capacity by boosting human intelligence and introspection through brain-computer interfaces, rather than tr…
A team at Redwood Research developed a proof-of-concept pipeline that transforms benign Claude Code transcripts into synthetic sabotage trajectories for automated red-teaming of AI monitors. The appro…