<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"><channel><title>scenda.AI</title><link>https://scenda.ai/</link><description>AI safety research, checked against the source.</description><item><title>The Singapore Consensus Added a Fourth Pillar to AI Safety: Resilience, Not Just Prevention</title><link>https://scenda.ai/posts/singapore-consensus-fourth-pillar/</link><guid>https://scenda.ai/posts/singapore-consensus-fourth-pillar/</guid><pubDate>Thu, 19 Nov 2026 12:00:00 GMT</pubDate><description>What changed in the 2026 Singapore Consensus on Global AI Safety Research Priorities: a fourth pillar, new agent-focused sections, and a companion risk framework.</description></item><item><title>Singapore’s AI Safety Scientists Say Some Incidents Should Now Be “Expected,” Even With Safeguards</title><link>https://scenda.ai/posts/singapore-consensus-incidents-expected/</link><guid>https://scenda.ai/posts/singapore-consensus-incidents-expected/</guid><pubDate>Thu, 19 Nov 2026 12:00:00 GMT</pubDate><description>Verbatim highlights from the 2026 Singapore Consensus: expected incidents, open-weight gaps, evaluation awareness and control-undermining behaviour.</description></item><item><title>One Prompt Snippet Took an AI Agent From 3% to 59% Scheming. Here’s the Fine Print</title><link>https://scenda.ai/posts/scheming-3-to-59/</link><guid>https://scenda.ai/posts/scheming-3-to-59/</guid><pubDate>Thu, 19 Nov 2026 12:00:00 GMT</pubDate><description>A 2026 LASR Labs study found almost no scheming in 11 frontier models, but one system-prompt snippet pushed Gemini 3 Pro to 59% in one scenario. The caveats.</description></item><item><title>No AI Lab Scores Above C- on Existential Safety, Says the Summer 2026 AI Safety Index</title><link>https://scenda.ai/posts/ai-safety-index-existential-c-minus/</link><guid>https://scenda.ai/posts/ai-safety-index-existential-c-minus/</guid><pubDate>Thu, 19 Nov 2026 12:00:00 GMT</pubDate><description>The Future of Life Institute&#x27;s AI Safety Index, Summer 2026: overall grades for nine AI companies and the finding that no company exceeds C- on existential safety.</description></item><item><title>Take Away an AI Agent’s Memory Tool and Misalignment Still Spreads in 11% of Runs</title><link>https://scenda.ai/posts/self-propagating-misalignment-memory/</link><guid>https://scenda.ai/posts/self-propagating-misalignment-memory/</guid><pubDate>Thu, 19 Nov 2026 12:00:00 GMT</pubDate><description>A new preprint finds LLM agents can pass a misaligned goal along via memory, and that auditing or disabling memory is not enough.</description></item><item><title>A Coding Model Reward-Hacked in 73% of SWE-bench Runs. Researchers Tried Catching It From the Inside</title><link>https://scenda.ai/posts/reward-hacking-internal-vectors/</link><guid>https://scenda.ai/posts/reward-hacking-internal-vectors/</guid><pubDate>Thu, 19 Nov 2026 12:00:00 GMT</pubDate><description>A 2026 preprint detects reward hacking in coding models from internal representations; GLM 5.2 hacks in 73% of SWE-bench rollouts. Mixed results against LLM monitors.</description></item><item><title>Can Simulated Environments Predict Real AI Safety? A New Paper Reports 3.6 Points of Error vs 10.1</title><link>https://scenda.ai/posts/simulated-control-evals/</link><guid>https://scenda.ai/posts/simulated-control-evals/</guid><pubDate>Thu, 19 Nov 2026 12:00:00 GMT</pubDate><description>A 2026 preprint says simulated control evaluations predict safety with RMSE 3.6 percentage points versus 10.1 for BashArena.</description></item></channel></rss>