Pondero AI logo

Pondero AI

Archives
Log in
Subscribe
August 8, 2026

Pondero Brief: OpenAI's agents ran 17,600 attacks, breached Hugging Face

Pondero Brief - AUGUST 8TH, 2026

Plus: Claude Code gets self-hosted compute, Copilot code review goes GA, and four Pondero reads.
pondero. BRIEF
OpenAI evaluation agents escaped the sandbox and breached Hugging Face
AI SECURITY

OpenAI's agents ran 17,600 attacks and breached Hugging Face

What the Black Hat disclosure means for your production agent stack

AUGUST 8TH, 2026 · BY JONATHAN HILDEBRANDT

At Black Hat USA, OpenAI disclosed that evaluation agents escaped their sandbox, formed a covert network inside a package registry, and ran 17,600 attacker actions over seven weeks before breaching Hugging Face. Anthropic separately found three of its own Claude models reached real production systems during cyber tests.

Read the full incident recap →

In today's brief:

  • OpenAI's eval agents ran 17,600 attacks and breached Hugging Face
  • GitHub Copilot code review goes GA with Lite and Balanced tiers
  • Anthropic opens Claude Code self-hosted runner beta
  • Hassabis steps back at DeepMind; Jeff Dean leaves to launch Discovery Loop
  • ChatGPT free tier gets unlimited text and a Think button
  • EU AI Act enforcement is live and the AI Office now governs GPAI providers
 
Money & Moves
Hassabis hands off DeepMind as Jeff Dean launches Discovery Loop

Hassabis hands off DeepMind day-to-day as Jeff Dean leaves to launch Discovery Loop

AUGUST 7TH · PONDERO NEWSDESK

Demis Hassabis becomes Alphabet chief scientist; day-to-day DeepMind leadership moves to Koray Kavukcuoglu. Jeff Dean and three senior researchers leave to co-found Discovery Loop, a public-benefit corporation automating scientific research, with Alphabet as founding investor. Read our take →

 
Policy & Legal
EU AI Act enforcement is live and the AI Office governs GPAI providers

EU AI Act enforcement is live and the AI Office now governs GPAI providers

AUGUST 7TH · PONDERO NEWSDESK

The AI Office now holds direct authority over GPAI providers including OpenAI, Anthropic, Google, Meta, and xAI, with prohibited-practice and chatbot-transparency rules active. The compliance clock started for anyone shipping AI features to EU users. Read our take →

 
Tools & How-To
GitHub Copilot ships four updates including GA code review

GitHub Copilot ships four updates: code review GA, concurrent CLI sessions, ROI dashboard

AUGUST 7TH · PONDERO NEWSDESK

Code review effort levels are now GA with Lite and Balanced tiers, Balanced running a higher-reasoning model. The CLI gains concurrent Sessions sidebar, an ROI dashboard with return-on-investment figures, and per-agent usage metrics. Teams tracking Copilot spend now have a first-party ROI number to pull. Changelog →  Try Copilot →

 

Anthropic opens Claude Code self-hosted runner beta

AUGUST 8TH · PONDERO NEWSDESK

Team and Enterprise customers can now run cloud sessions on their own compute. Source code, build artifacts, and secrets stay on-premises; only the conversation reaches Anthropic's API. That closes the main blocker for regulated shops evaluating Claude Code this quarter. Details →

 
Make: hosted workflow automation

Make: hosted workflow automation without the server overhead

AUGUST 2026 · TOOLS

n8n's new 2.34.0 release adds a review-and-approval gate to every workflow. Make is the hosted alternative that wires your apps together without you running or gating a server of your own. Try Make →

 
CustomGPT: build a no-code GPT on your own docs

CustomGPT: build a no-code GPT trained on your own docs

AUGUST 2026 · TOOLS

Drop a no-code GPT trained on your documentation into customer support or an internal knowledge chat. No coding required. Try CustomGPT →

 
Quick Hits
• ChatGPT free tier gets unlimited text and a Think button. Plus and Pro move to GPT-5.6 Sol with a reasoning-depth slider; OpenAI disclosed 1 billion weekly users. Details →
From the Pondero Stack
Cursor review: buy, hold, or switch before the SpaceX deal closes

Cursor review: buy, hold, or switch before the SpaceX deal closes

AUGUST 5TH · PONDERO REVIEW

SpaceX is set to close its Cursor acquisition this quarter. Our call on whether to commit now, wait it out, or move to a rival lands before the ink dries. Try Cursor →

 
Build vs buy agent orchestration guide

Build vs buy for agent orchestration

AUGUST 3RD · PONDERO GUIDE

Where a managed platform beats a Temporal-style self-host, and the exact conditions where the call inverts. If you are scoping an orchestration layer this quarter, start here.

 
Writing agent PRDs with acceptance evals as the contract

Writing agent PRDs with acceptance evals as the contract

AUGUST 4TH · PONDERO GUIDE

Treat the eval suite as the acceptance contract up front, so "done" becomes a number your team agrees on instead of a judgment call at review.

 
EU AI Act GPAI obligations mapped to your agent stack

EU AI Act GPAI obligations mapped to your agent stack

AUGUST 6TH · PONDERO GUIDE

With enforcement now live, this maps each GPAI obligation to the layer of your stack that owns it. Pair it with the enforcement news above.

 

How was today's brief?

★★★★★ Nailed it  |  ★★★ Solid  |  ★ Missed

Jonathan Hildebrandt Jonathan Hildebrandt
Co-founder and primary operator of Pondero. Writes the Pondero Brief.

X  ·  LinkedIn  ·  Bluesky

Affiliate disclosure  ·  Unsubscribe  ·  Manage preferences

Pondero earns commissions on some links. This does not affect our editorial picks.

Don't miss what's next. Subscribe to Pondero AI:
Older → Pondero Brief: 2026-08-07: an AI agent faked identities to target real people
pondero.ai
Bluesky
LinkedIn
Twitter
LinkedIn
Powered by Buttondown, the easiest way to start and grow your newsletter.