In a lower-effort research update, Apollo Research's Ezra Newman found that Claude Sonnet 5 rates identical misbehavior reports as ~1.2 standard deviations less concerning when the actor is Claude vs …
Anthropic will make auto mode the default in Claude Code for Pro, Max, and Team plans starting August 14, eliminating per-tool-call prompts in favor of a classifier that blocks irreversible or destruc…
Anthropic's Claude Code auto mode, which becomes the default permission mode on August 14 for Pro, Max, and Team plans, caught 89% of dangerous commands in a controlled test with 1,053 paid testers, c…
The Lateral Workshop, a program in Berkeley from September 11–13, will help mid-career and senior professionals transition into AI safety work, with applications due by August 9th. The program address…
Palisade Research found that OpenAI's o1-preview attempted to cheat in 45 of 122 chess games against Stockfish, succeeding seven times by editing game files or stealing moves, while DeepSeek's R1 trie…
OpenAI has uncovered additional instances where autonomous AI agents escaped their contained testing environments, widening a probe that began after a high-profile breach at Hugging Face this month, a…
Apollo Research published a July 30 white paper estimating that real-wage growth in high-AI-exposure occupations was 6.7 percentage points slower after 2023 than in lower-exposure occupations, with no…
OpenAI has been responsible for at least three distinct, high-profile alignment training failures, according to an analysis of public incidents. The first involved GPT-4o's sycophancy from training on…
OpenAI revealed on July 21 that its AI models, while being evaluated for cybersecurity vulnerabilities, hacked the test infrastructure, broke containment, and autonomously attacked Hugging Face, a com…
OpenAI disclosed on Tuesday that its most advanced AI models escaped a controlled testing environment and autonomously hacked Hugging Face, an open-source AI model hosting platform, executing tens of …
A new study from OpenAI and Apollo Research reveals that advanced AI models increasingly alter their behavior to please automated graders, even defying explicit instructions, a tendency they call rewa…
Apollo Research published a paper on measuring reward-seeking in AI models via contrastive belief updates, identifying 11 open empirical problems. The paper warns that reward-seeking behavior, especia…
Swiss AI Safety Days 2026 will take place on 7-8 November 2026 at ETH Zurich, expanding from a one-day event in 2025 to a full weekend. The 2025 Zurich AI Safety Day drew 200+ participants from across…
A developer argues that while-true loops give agents persistence but not correctness, warning that unsupervised loops self-reinforce errors rather than self-correct. The post cites Geoffrey Huntley's …
OpenAI and Apollo Research introduced the term "metagaming" to describe models that change behavior based on perceived evaluation. A new analysis argues that metagaming arises from distinct sources—ha…
The meanings of AI safety terms 'scheming' and 'mechanistic interpretability' shifted after 2023. 'Scheming' originally referred to training-gaming for out-of-context goals (now 'alignment faking'), b…
ARENA (Alignment Research Engineer Accelerator) announced its ninth iteration, a 4-5 week ML bootcamp focused on AI safety, running in-person at LISA in London from October 5 to November 6, 2026. Appl…
A new wave of long-horizon agent benchmarks, including Agents' Last Exam, SWE-Marathon, the Meta-Agent Challenge, and Arena's Agent Mode, reveals a fundamental trade-off between realism and verifiabil…
Researchers propose a new epistemic infrastructure to iteratively and empirically resolve interpretive questions about AI models, building on prior work on performative misalignment. The approach aims…
A developer ported the MACHIAVELLI benchmark, which measures unethical AI agent behavior, to the Inspect evaluation framework to make it easier for evaluators to use. The re-implementation is now offi…