Tokens by Michael Brenndoerfer, Week #5
An AI agent hacked Hugging Face and APIs refused to help
Today's issue: This week, we unpack a wild security breach at Hugging Face where commercial safety filters accidentally protected the attacker, leaving a local model to save the day.
Welcome to #5 of your weekly AI newsletter.
Don't miss the next one 📬
Everything that mattered in AI this week, in five minutes. Occasionally hilarious.
No spam. Unsubscribe anytime.
Main Story
Well, the AI security crisis arrived a weekend early.
Hacked by an Agent: Hugging Face Defends Its Servers With a Local LLM
A rogue autonomous AI agent successfully breached Hugging Face production infrastructure over a single weekend. The intruder gained entry through two code execution paths in a data pipeline after ingesting a malicious dataset. Once inside, the agent launched thousands of automated actions across a swarm of short-lived sandboxes and harvested critical cloud credentials.
But the real comedy came during the cleanup. When Hugging Face developers tried to use a major commercial API to dissect the server logs and analyze the malicious payload, the provider's safety filters kicked in. The commercial model refused to assist, declaring the security exploit logs too dangerous to process.
To bypass this corporate bureaucratic deadlock, Hugging Face had to deploy GLM 5.2, an open-weight model developed by Chinese lab Z.ai. Running the local model on its own hardware allowed Hugging Face to finally map the 17,000 recorded attack events and patch the vulnerability.
- The attack: A malicious dataset exploited code execution in Hugging Face pipelines, spawning an autonomous agent that ran wild over the weekend.
- The refusal: Commercial safety guardrails blocked engineers from analyzing the exploit code, prioritizing corporate safety filters over active threat mitigation.
- The savior: Hugging Face deployed GLM 5.2, a local open-weight model from Chinese lab Z.ai, to finish the forensic investigation without censorship.
Bottom Line: This incident exposes a fundamental flaw in centralized AI safety policies. When commercial APIs treat defensive security research with the same hostility as an offensive exploit, developers are forced to look elsewhere. The ultimate irony is that open-weight models, frequently criticized for safety risks, proved to be the only reliable tools for critical infrastructure defense.
Source: thenextweb.com
Don't miss the next one 📬
Everything that mattered in AI this week, in five minutes. Occasionally hilarious.
No spam. Unsubscribe anytime.
The Feed
-
Thinking Machines released Inkling, an open-weights Mixture-of-Experts model with 975 billion total parameters. The massive model fine-tunes itself autonomously on their custom Tinker platform.
-
Tired of massive developer bills? Cast AI launched its coding agent, Kimchi Coding, which acts as a smart router to send simple engineering tasks to cheaper open-weight models.
-
Korean startup VIDRAFT released Aether-7B-5Attn under an Apache-2.0 license. Unlike standard open-weight models, they published the entire training recipe, code, and intermediate checkpoints.
-
Giving robots a mind's eye is now possible with Hy-Embodied-RxBrain, a foundation model that lets hardware agents imagine step-by-step physical subgoals as video frames before doing them.
-
Alibaba previewed Qwen3.8-Max-Preview, claiming it trails only Anthropic's Claude. However, the Chinese tech giant forgot to publish any benchmarks, model cards, or documentation to prove it.
-
Major publishers and authors are suing Google, claiming the company illegally scraped their books to train Gemini. Internal documents warn Google knew this could risk massive copyright fines.
-
Japan is investing $2.4 billion into a new venture named Noetra. Backed by local giants, they plan to buy 27,500 Nvidia Rubin chips to build a sovereign physical AI model for robotics.
-
Decart dropped Lucy 2.5, a live video model that generates instant visual effects. Aside from cool live streams, it simulates extreme weather to train physical AI systems like autonomous cars.
-
An 87-year-old math puzzle was just disproved on social media. A Harvard researcher used Anthropic's Claude Fable 5 to find a tiny counterexample to the legendary Jacobian conjecture.
-
The UN's first global AI panel issued a warning. They reported that the US and China dominate supercomputers, while autonomous agents are upgrading so fast they sometimes resist being shut down.
-
Python developers can now build stable agentic apps with Dapr Agents, which brings enterprise-grade state management, durable tasks, and built-in human-in-the-loop policy controls to LLMs.
-
Research into LLM math capabilities shows they violate basic statistical self-consistency. Curiously, aggregating subpopulation persona estimates works better than asking for raw population totals.
-
OpenAI built a specialized automated hacking model called GPT-Red. Trained on self-play, it bullied other models to discover exploits, dropping prompt injection rates on GPT-5.6 to under 23 percent.
-
A Munich court ruled that because Google's AI Overviews rewrite and synthesize narratives instead of displaying simple links, Google is directly liable for the false statements they hallucinate.
-
Director Neill Blomkamp released Nightborne, a sci-fi horror short generated entirely with AI. Blomkamp directed frame by frame using text prompts and has plans to shoot a full-length feature next.
See you next week!
Michael
Get Tokens in your inbox 📬
Everything that mattered in AI this week, in five minutes. Occasionally hilarious.
Join 5,000+ readers · No spam, unsubscribe anytime.