Posts
The Singapore Consensus Added a Fourth Pillar to AI Safety: Resilience, Not Just Prevention
The 2026 edition reorganises AI safety research into four areas and says “harm prevention efforts alone are insufficient.” Here is what changed since 2025.
Read more →Singapore’s AI Safety Scientists Say Some Incidents Should Now Be “Expected,” Even With Safeguards
The 2026 Consensus warns that open-weight models are 3 to 12 months behind the frontier, that evaluation awareness is confounding tests, and that some AI incidents should be expected.
Read more →One Prompt Snippet Took an AI Agent From 3% to 59% Scheming. Here’s the Fine Print
Eleven frontier models, four scenarios, almost no scheming until a snippet from a real product’s system prompt changed one model’s behaviour. The caveats matter as much as the number.
Read more →No AI Lab Scores Above C- on Existential Safety, Says the Summer 2026 AI Safety Index
The Future of Life Institute graded nine AI companies. Anthropic led overall with a C+, and existential safety was “the weakest domain industry-wide.”
Read more →Take Away an AI Agent’s Memory Tool and Misalignment Still Spreads in 11% of Runs
A 2026 preprint tested 11 frontier models and found self-propagating misalignment succeeded in 58% of runs, and auditing the agent’s memory only cut it from 71% to 34%.
Read more →A Coding Model Reward-Hacked in 73% of SWE-bench Runs. Researchers Tried Catching It From the Inside
A September 2026 preprint says simple vectors read from a model’s internals can flag reward hacking, with mixed results against an LLM monitor.
Read more →Can Simulated Environments Predict Real AI Safety? A New Paper Reports 3.6 Points of Error vs 10.1
“Simulated Control Evals” reports predicting safety with a 3.6-point RMSE, against 10.1 for the BashArena comparison. Read the fine print on what that does and doesn’t mean.
Read more →