Microsoft Research: editing under 10 neurons can backdoor an AI agent
A trigger string makes the model run an attacker command with the agent's access; Microsoft has not seen real-world use.
1
Stories per day · 30 days
23 RSS
Relatedmixture-of-experts22huggingface19multimodal16nvidia13benchmarks11nemotron11agents9hugging-face9model-release9meta8moe8qwen8deepseek7kimi7minimax7mistral7muse-spark-open7on-device7
329 stories, newest first · one card per story
Stories tagged #open-weights
MBZUAI’s IFM releases open 29B PAN 2 world model
Apache 2.0 weights and code ship; IFM reports 34% full-plan completion, ahead of MiniMax-H3 on multi-step tasks.
1
Sixteen LLMs find new TIS-100 score frontiers on 18 of 55 puzzles
Opus 5.5 and GPT-6 Astra each solved all 55; Opus found 72 new frontier points to Astra's 32.
1
Anthropic IPO Valuation Slumps to $80B
AI market rout deepens with valuations revised downward as Anthropic targets $100B run rate.
1
NaiveAI ships open-weights Naive-N0.5-Flash for fast coding
A small Latent Space test saw high server-side throughput and very long hard-question traces, without grading accuracy.
1
Intezer agents rebuild Slack, Finder and TinyTurla-NG from binaries
The TinyTurla-NG rebuild scored 92.8% BinDiff similarity, and the Slack copy gained local AI autocomplete.
OpenAI and Anthropic limit accounts used for AI reverse engineering
Users report Claude and Codex blocks while modifying owned devices; an OpenAI engineer says the blocks are accidental.
1
Sable proves bit-exact quantized LLM inference on one GPU
Qwen2.5-32B takes 92.5 seconds for 4,096 tokens; proofs are not zero-knowledge.
External monitors outperform prompt guardrails against agent injection
On two Qwen2.5 models, a task-bound monitor admitted 1.3% and 7.7% of attacks after models proposed the attacker's call in 64% and 84%.
ConwayResearch publishes Underdog Woof, a 40.6 MB English TTS model
The model card claims 1.42% word error rate for Mira and CPU-only synthesis with numpy.
CallCleaner distills SAM-Audio into a 19M bioacoustic denoiser
Authors say the CPU model runs about 90 times faster than real time and leaves 11 dB less noise than Biodenoising.
Nous Research Releases Arete-27B Fine-Tuned Qwen Model
Arete-27B is a 27B open-weight model reducing propaganda answers from 22% to 0.9% on an 803-prompt eval.
1
Tencent releases Hy4 with vision and computer use; weights come later
Hy4 is API-only for now, scores 35 on Artificial Analysis, and trails Kimi K3.1 Pro by a wide margin.
AI Industry News: Model Releases and Executive Moves
Moonshot, Ricursive, Cognition release new open-weight AI models while several executives switch between companies.
1
Video Rebirth publishes HyperFlow H3 Flash, a compressed MiniMax-H3
The 13.7B–13.9B model makes 5.2–15 second video with stereo audio in 4–8 steps, under a restricted MiniMax license.
Micius-Flash, a 27B open engineering model, is released
It averages 64.2% on hardware code benchmarks and 62.4% on a semiconductor agent benchmark, above a same-size baseline.
MiniMax releases open-weight MiniMax-H3-Image for generation and editing
The checkpoint outputs about 16-megapixel images in 6 steps and accepts up to 16 reference images.
Moonshot AI releases Kimi K3.1 Pro open-weight multimodal model
Kimi K3.1 Pro is an open-weight MoE model with 1M token context and 104B activated parameters.
1
Mistral open-sources Large 4 weights despite cyberattack uplift warning
Downloaded weights sit outside API guardrail updates, and Mistral has not said whether Large 4 meets the EU systemic-risk threshold.
1
Essay says SOCs are unready for autonomous offensive agents
It cites OpenAI agents attacking Hugging Face and a GLM-5.3 browser exploit built in a day.
1
Mistral Large 4 ships as open weights for coding and cyber
The 1.05T model has 52B active parameters and a 1M context; Mistral's ranking claims are narrower than independent scores.
4
Essay argues cheap AI agents will make cyberattacks routine
It cites an authorized NSA Mythos test, a claimed Cairn campaign, and slower patching times.
1
Pika releases Pika H3 Flex open weights for video and audio
Local inference outputs native 480p or 768p with stereo audio; the hosted API adds up to 4K.
3
Mistral releases Large 4 open weights, nicknamed Le Chonk
The trillion-parameter model leads some open coding scores but trails top Chinese models and Claude overall.
1