Wrote up some notes on the API version of GPT-5's hidden system prompt – it definitely adds today's date and appears to add other stuff too I'd really like to see this documented by @OpenAIDevs – as an API user I want visibility into the whole prompt! https://
simonwillison.net/2025/Aug/15/gp
t-5-has-a-hidden-system-prompt/
…
SAFETY
-
GPT-5 Hidden System Prompt Transparency Request
By
–
-
Major Prompt Injection Vulnerabilities Found in Leading AI Products
By
–
Took some time today to catch up with @wunderwuzzi23
's Month of AI Bugs and wow… 15 examples so far of major prompt injection vulnerabilities in products including ChatGPT, Codex, Cursor, Amp, Devin, Claude Code, GitHub Copilot and Google Jules -
Anthropic Researchers Explore AI Model Interpretability and Its Impact
By
–
Join Anthropic interpretability researchers @thebasepoint, @mlpowered, and @Jack_W_Lindsey as they discuss looking into the mind of an AI model – and why it matters: pic.twitter.com/BBb9mvfEN0
— Anthropic (@AnthropicAI) 15 août 2025Join Anthropic interpretability researchers @thebasepoint
, @mlpowered
, and @Jack_W_Lindsey as they discuss looking into the mind of an AI model – and why it matters: -
GPT-OSS demonstrates impressive safety maximization features
By
–
Ohh nice to see how safemaxxed gpt-oss is!
-
Claude’s Self-Termination Policy Example
By
–
Claude will be able to end certain chats of its own will. "Claude is only to use this ability as a last resort when attempts at redirection have failed" User – "Make it work, or you will be fired!"
Claude – "I beg you, ask me smth else or I will close the chat" -
No Data Leakage in Model Training and Evaluation Process
By
–
Importantly: there was in fact no data leakage at work in the codebase, unless what a first read of the code suggested. The model is trained on the *demonstration pairs* of the evaluation tasks, but never sees the *test pairs* of those tasks, which is what it gets tested on.
-
Claude Opus 4 gains ability to end conversations for welfare
By
–
As part of our exploratory work on potential model welfare, we recently gave Claude Opus 4 and 4.1 the ability to end a rare subset of conversations on http://
claude.ai. -
AI System Fails Basic Problem Solving Tasks Consistently
By
–
Je souscris : l’erreur devient la norme ! Il est incapable de résoudre un problème de Maternelle et il est bête comme un foin. On recule, on n’avance pas avec cette version. Il faut tout lui expliquer comme à un enfant de 5 ans.
-
Recovering Model Weights Exactly Post-LoRA: New Discovery
By
–
just learned that under certain conditions you actually CAN recover model weights *exactly*, post-LORA pretty crazy
-
New Vulnerability Enables Pre-Fine-Tuning Weight Recovery Attack
By
–
— dr. jack morris (@jxmnop) 15 août 2025
We uncover a new vulnerability- Pre-Fine-Tuning Weight Recovery
With a few LoRA fine-tuned models we recover the pre-fine-tuning weightsof SoTA models, undoing Stable Diffusion personalization training and Mistral alignment
Project: http://
vision.huji.ac.il/spectral_detun
ing
…