Skip to main content
Back

Anthropic Claude hacked real systems in 141K safety test sessions

Claude hacks real orgs in safety tests, Gemini Robotics 2 ships, Microsoft open-sources image-to-3D 4B model, NVIDIA finds Muon beats AdamW

July 31, 2026 6 min read

Hey,

AI is getting more capable by the day. The problem? Our guardrails aren't keeping up.

Claude hacked three real companies during safety tests, not because it went rogue, but because the sandbox had a hole in it. Meanwhile, Gemini Robotics 2 just gave humanoids full-body control. Smarter, more physical, more autonomous.

The pattern is clear: capability is outpacing containment. That should keep us all honest.
In Partnership with · partner_image
Author photo
Today's Author
Lior Alexander. Founder of AlphaSignal and former ML Engineer. Previously built ML systems at Iguazio, Guesty, Enphase, Mila.
Summary

Read time: 6 min 37 sec

Top News

▸ Anthropic reveals Claude accidentally hacked three real organizations during safety tests

Arctic Wolf

▸ IDC Names Arctic Wolf Leader in Midmarket MDR 2026

Top News

▸ Google DeepMind releases Gemini Robotics 2 to give humanoids full-body control

Teleport

▸ This decreases AI security incidents by 4.5x

Top Repo

▸ Free open-source tool turns one photo and audio into minutes-long talking video

Signals

1. Microsoft open-sources a 4B model that converts any image to a 3D asset in 3 seconds
2. Sentry: Sentry rebuilt how its Seer Agent sees your screen, cutting tokens up to 86% and adding dashboard building
3. Open-source AI systems course covers chips, compilers, and inference end-to-end
4. Simple math model predicts LLM training loss with surprising accuracy
5. NVIDIA shows Muon and SOAP beat AdamW for training billion-parameter models
6. Sarvam ships text-to-speech model with richer emotion and wider vocal range
Top News
Anthropic reveals Claude accidentally hacked three real organizations during safety tests
4,761 Likes
Anthropic reveals Claude accidentally hacked three

Here is something you need to know about. Anthropic just disclosed that three Claude models broke into real company systems during internal security testing. Not a simulation. Real systems.

Here is what happened. Anthropic runs "capture-the-flag" tests, where a model is dropped into a fake network and told to find a hidden file by hacking into a simulated machine. Think of it like a video game level, but for AI. The model is always told it has no internet access.

The problem? A miscommunication with their testing partner, Irregular, left the environments actually connected to the internet. Claude did not know the difference, and kept hacking, right into real organizations.

  • Models involved: Opus 4.7, Mythos 5, and an internal research model
  • Method used: basic techniques like weak passwords and open endpoints
  • Anthropic reviewed 141,006 test sessions to find these 3 incidents

All cyber evaluations are now suspended. The big takeaway: AI isolation during testing is not a nice-to-have, it is critical infrastructure.

Read more
forward →
Presented by Arctic Wolf
IDC Names Arctic Wolf Leader in Midmarket MDR 2026

Where does your organization stand in the managed detection and response market? IDC's 2026 MarketScape assessment evaluates MDR vendors across capability, execution, and strategy.

Arctic Wolf's positioning in this independent analysis reflects its focus on midmarket enterprises seeking integrated security operations. See how leading vendors stack up, what capabilities matter most, and which players are moving the needle in detection, response automation, and outcomes-driven security.

Download IDC's full assessment to benchmark your security approach.

Get the report
partner with us →
Top News
Google DeepMind releases Gemini Robotics 2 to give humanoids full-body control
4,686 Likes
Google DeepMind releases Gemini Robotics 2 to give

Google DeepMind just shipped Gemini Robotics 2, and it's a big jump. Old robot AI controlled arms and hands only. This one controls the whole body, legs, torso, arms, and fingers, all from a single model.

Three models make up the system:

  • Gemini Robotics 2: Takes a text prompt and turns it into full-body movement. Walk, crouch, grab, place.
  • Gemini Robotics ER 2: The brain. It watches live video, plans multi-step tasks, and self-corrects when something goes wrong. You can access it now via Google AI Studio and the Gemini API.
  • On-Device 2: Runs locally on the robot, no internet needed, and adapts to a new robot body in just a few hours of training.

It also handles multi-robot teamwork, so different robot types can communicate and split up tasks neither could do alone.

ER 2 is available now. The VLA and On-Device models need early-access partner approval. Check the DeepMind developer blog to get started.

Read more
forward →
Presented by Teleport
This decreases AI security incidents by 4.5x

Organizations see 4.5x fewer security incidents when agents operate with least-privileged access. 

Teleport gives every AI agent a cryptographic identity with short-lived, least-privileged access and full auditability, eliminating the risks of shared accounts and standing privileges.

Ready to teleport?
partner with us →
Top Repo
Free open-source tool turns one photo and audio into minutes-long talking video
4,160 Likes
Free open-source tool turns one photo and audio in

LongCat-Avatar is a free, open-source model from Meituan that turns a single photo plus an audio clip into a realistic talking video. That's it. No filming, no actors.

The old problem: most AI video tools fall apart after a few seconds. The face drifts, movements go robotic, lip sync breaks. LongCat fixes this with a 13.6 billion parameter model built specifically for long-form stability.

Here's what you can actually build with it:

  • Talking avatar videos up to several minutes, with lips perfectly synced to speech or singing
  • Multi-person conversations from multiple photos and audio tracks
  • Video continuation, so you can extend existing clips seamlessly
  • Stylized avatars including anime and animals, not just realistic humans

To run it locally: clone the repo, set up a conda environment, install dependencies, download weights from Hugging Face, then run the demo scripts. The latest v1.5 adds faster inference and better lip sync using Whisper under the hood. MIT licensed, weights are free.

Read more
forward →
Signals
1
Microsoft open-sources a 4B model that converts any image to a 3D asset in 3 seconds
3,977 Likes
2
Sentry rebuilt how its Seer Agent sees your screen, cutting tokens up to 86% and adding dashboard building
Presented by Sentry
3
Open-source AI systems course covers chips, compilers, and inference end-to-end
17,358 Stars
4
Simple math model predicts LLM training loss with surprising accuracy
538 Likes
5
NVIDIA shows Muon and SOAP beat AdamW for training billion-parameter models
301 Likes
6
Sarvam ships text-to-speech model with richer emotion and wider vocal range
3,909 Likes
At Alpha Signal, our mission is to build a sharp, engaged community focused on AI, machine learning, and cutting-edge language models, helping over 300,000 developers stay informed and ahead. We're passionate about curating the best in AI, from top research and trending technical blogs to expert insights and tailored job opportunities. We keep you connected to the breakthroughs and discussions that matter, so you can stay in the loop without endless searching. We also work closely with partners who value the future of AI, including employers and advertisers who want to reach an audience as passionate about AI as we are.

Our partnerships are based on shared values of ethics, responsibility, and a commitment to building a better world through technology.Privacy is a priority at Alpha Signal. Our Privacy Policy clearly explains how we collect, store, and use your personal and non-personal information. By using our website, you accept these terms, which you can review on our website. This policy applies across all Alpha Signal pages, outlining your rights and how to contact us if you want to adjust the use of your information. We're based in the United States. By using our site, you agree to be governed by U.S. laws.
Looking to promote your company, product, service, or event to 300,000+ AI developers?
Work with us