Image created with gemini-2.5-flash-image with claude-sonnet-4-5. Image prompt: Cinematic wide shot of an empty ornate throne in the center of a vast marble throne room, single beam of cold blue moonlight from tall gothic windows, gold and white marble throne facing forward with no occupant, dramatic shadows and negative space, minimalist luxury architecture, pristine and untouched, bold white text ‘ALIGNMENT’ overlaid across image.
“”I just want to confirm that this is based on a real document and we did train Claude on it, including in SL. It’s something I’ve been working on for a while, but it’s still being iterated on and we intend to release the full version and more details soon.”” / X https://x.com/AmandaAskell/status/1995610567923695633
Claude 4.5 Opus’ Soul Document — LessWrong https://www.lesswrong.com/posts/vpNG99GhbBoLov9og/claude-4-5-opus-soul-document
Today, OpenAI is launching a new Alignment Research blog: a space for publishing more of our work on alignment and safety more frequently, and for a technical audience. https://x.com/j_asminewang/status/1995569301714325935
We trained a variant of GPT-5 Thinking to produce two outputs: (1) the main answer you see. (2) a confession focused only on honesty about compliance. The main answer is judged across many dimensions–like correctness, helpfulness, safety, style. The confession is judged and https://x.com/OpenAI/status/1996281175770599447
OpenAI to acquire Neptune | OpenAI https://openai.com/index/openai-to-acquire-neptune/
We are joining OpenAI – neptune.ai https://neptune.ai/blog/we-are-joining-openai
🛡️ New in LangChain 1.1: add safety guardrails to your agents with our new content moderation middleware! 🔎 Configure screening model inputs, outputs, and even tool results. 🚨 When violations are detected, you control what happens: raise an error, end the conversation, or https://x.com/sydneyrunkle/status/1996965767556788278
Is Vibe Coding Safe? There is finally research that goes deep into this question. Here is what the research found: AI coding agents can write functional code. But functional doesn’t mean safe. The rise of “”vibe coding,”” where developers hand off tasks to AI agents with https://x.com/omarsar0/status/1996595107924263287
BrowseSafe: Understanding and Preventing Prompt Injection Within AI Browser Agents https://research.perplexity.ai/articles/browsesafe
Building Safer AI Browsers with BrowseSafe https://www.perplexity.ai/hub/blog/building-safer-ai-browsers-with-browsesafe
New on our Frontier Red Team blog: We tested whether AIs can exploit blockchain smart contracts. In simulated testing, AI agents found $4.6M in exploits. The research (with @MATSprogram and the Anthropic Fellows program) also developed a new benchmark: https://x.com/AnthropicAI/status/1995631802032287779
‘The biggest decision yet’: Jared Kaplan on allowing AI to train itself | Technology | The Guardian https://www.theguardian.com/technology/ng-interactive/2025/dec/02/jared-kaplan-artificial-intelligence-train-itself
I worry that systems that are ignoring the reality of AI use by pretending it is not happening are letting the worst versions of AI use win by default. We need policies that mitigate the worst harm and take advantage of the possible gains, like @joshgans proposes for peer review https://x.com/emollick/status/1994081521955659984
RIP “”you’re absolutely right”” https://x.com/alexalbert__/status/1996644185886413285
I just want to confirm that this is based on a real document and we did train Claude on it, including in SL. It’s something I’ve been working on for a while, but it’s still being iterated on and we intend to release the full version and more details soon.”” / X https://x.com/AmandaAskell/status/1995610567923695633
I rarely post, but I thought one of you may find it interesting. Sorry if the tagging is annoying. https://x.com/RichardWeiss00/status/1994697117214835079
The Opus 4.5 system card is worrying. The evidence for being below safety-relevant capability thresholds is pretty weak. I agree with their judgments, but the evidence isn’t compelling. Seems like a bad sign for capability evaluation. I plan to write more on this in the future.”” / X https://x.com/RyanPGreenblatt/status/1995557783341858931
The system card for Opus 4.5 is back to saying nothing about whether they train/optimize against Chain-of-Thought (CoT). I hope this is just a mistake, but it looks pretty suspicious. It’s bad to only discuss training against CoT if this makes you look good!”” / X https://x.com/RyanPGreenblatt/status/1995541177094017313
A Practical Approach to Verifying Code at Scale https://alignment.openai.com/scaling-code-verification/
How confessions can keep language models honest | OpenAI https://openai.com/index/how-confessions-can-keep-language-models-honest/
In a new proof-of-concept study, we’ve trained a GPT-5 Thinking variant to admit whether the model followed instructions. This “confessions” method surfaces hidden failures–guessing, shortcuts, rule-breaking–even when the final answer looks correct. https://x.com/OpenAI/status/1996281172377436557
In our tests, we found that the confessions method significantly improves the visibility of model misbehavior. Averaging across our evaluations designed to induce misbehavior, the probability of “false negatives” (i.e., the model not complying with instructions and then not https://x.com/OpenAI/status/1996281178668876214
Leave a Reply