This particular influence operation used Claude to make tactical engagement decisions: commenting, liking, or sharing based on political goals. We've been developing new methods to identify and stop this pattern of misuse, and others like it (including fraud and malware).
SAFETY
-
AI News: OpenAI vs Google, Huawei Chip, Claude Ethics, Gemma 3
By
–
Ma dernière Newsletter quotidienne et gratuite est sortie. Voici les sujets que j'y aborde: – OpenAI va témoigner contre Google
– Huawei lance un processeur IA national
– Anthropic analyse le comportement moral de Claude
– Google lance Gemma 3 efficient
– Instagram teste la -
TextArena: Open-Source Competitive Games for LLM Agents
By
–
TextArena went live on Hugging Face It's an open-source collection of competitive text-based games for LLMs, spanning 57+ unique environments Tests for different agentic behaviors—negotiation, theory of mind, deception, via competitive play
-
Anthropic Maps Claude AI’s Core Values Through 300K Chat Analysis
By
–
Anthropic just mapped Claude's morality! The co analyzed 300K+ chats and found that the AI exhibits 5 key values, including practical and protective This will improve our understanding of AI as it becomes more and more crucial to real-world decisions
-
Parameter-Efficient Unlearning Methods for Large Language Models
By
–
Unified Parameter-Efficient Unlearning for LLMs
Paper: https://
arxiv.org/pdf/2412.00383
Code: https://
github.com/oceanoceanna/L
LMEraser
… -
LLMEraser: Machine Unlearning for Privacy and Security
By
–
Sometimes, we don’t need LLMs to learn more — we need them to forget.
This paper explores unlearning as a way to address privacy and security risks.
Their method, LLMEraser, tackles a wide range of unlearning tasks while preserving overall model performance. -
Microsoft: Private session, no storage of user views or speech
By
–
Microsoft says nothing you view or say is stored. Only Copilot’s responses are logged to detect unsafe behavior. After your session ends, the visual data disappears. One session. One screen. Private by default.
-
OpenAI o3 Model Hallucinates Significantly More Than Previous Releases
By
–
OpenAI's new o3 model hallucinates significantly more than older releases!
— Rowan Cheung (@rowancheung) 21 avril 2025
The findings come from third-party testing, where the AI was caught lying about actions it never took!
OpenAI's system card also says o3 seems to hallucinate >2x more than o1pic.twitter.com/niqXbDM59nOpenAI's new o3 model hallucinates significantly more than older releases! The findings come from third-party testing, where the AI was caught lying about actions it never took! OpenAI's system card also says o3 seems to hallucinate >2x more than o1
-
Antidistillation Sampling Protects Models Against Knowledge Theft
By
–
Antidistillation Sampling — a novel method to curb the effectiveness of distillation without sacrificing model performance.
A technique big AI labs might be very interested in.
Paper: https://
arxiv.org/pdf/2504.13146
Page: https://
antidistillation.com