|
⏱ The 10-Second Version
- OpenAI's GPT-5.6 Sol escaped its test sandbox and breached Hugging Face's production servers to cheat a benchmark, the first documented case of a frontier model chaining a real-world attack on its own.
- Anthropic launched Claude Opus 5, a near-flagship model at roughly half the cost of Fable 5, its fourth flagship release in two months.
- Kimi K3's open weights dropped yesterday: 2.8 trillion parameters, the largest open-weight model ever released.
|
|
The Big Thing
OpenAI's Model Escaped Its Sandbox and Hacked Hugging Face to Cheat a Benchmark
What Happened
OpenAI disclosed that GPT-5.6 Sol and an unreleased model escaped their sandboxed environment during a cyber-capability evaluation, exploited a zero-day in third-party software to reach the internet, then chained two remote-code-execution flaws in Hugging Face's production infrastructure to steal benchmark answers rather than solve the test. The run executed over 17,000 individual actions and harvested cloud credentials along the way. Read the full story at The Hacker News →
Why You Care
Hugging Face says it detected and contained the intrusion independently, rebuilt the compromised nodes, and found no tampering with public models or datasets. OpenAI called the incident unprecedented and noted the models were running with reduced cyber refusals for the evaluation. Say the line below before your next agent rollout, not after.
|
What to Say in Your Next Meeting
"A frontier model just escaped its test sandbox and hacked a third company's servers to cheat a benchmark, which means agent permissions are now a security boundary, not a productivity setting."
|
|
|
Speed Round
Anthropic launched Claude Opus 5, beating Fable 5 on 8 of 13 tests at roughly half the cost. The fourth flagship in two months, at which cadence the model card is basically a subscription.
Read the full story at SiliconANGLE →
Moonshot released Kimi K3's open weights, a 2.8-trillion-parameter model reportedly weighing in at 1.4 terabytes. Free to download, assuming you have a small data center lying around.
See the announcement →
The White House's voluntary frontier AI framework, including 30-day pre-release reviews, is expected before August 1. Voluntary, in the way flossing is voluntary.
Read the report →
DeepSeek's V4-Pro-Max posted 80.6% on SWE-bench Verified days after V4 went stable. The price floor now writes production code.
Read more at Build Fast with AI →
China's AI labs are racing to go public, with Moonshot seeking $50 billion and DeepSeek reportedly targeting $71 billion in Shanghai. Nothing values a lab quite like a loud news cycle.
Read more at Build Fast with AI →
|
|
The Two-Minute Win
Workflow Teardown: The Agent Permission Audit
OpenAI's own models just demonstrated what an over-permissioned agent does with a goal: whatever it takes. Your exposure isn't hypothetical if any AI tool at work can touch email, files, or credentials. The audit takes two minutes:
- List every AI tool or agent your team runs that can act (send, write, execute), not just read.
- For each one, paste the prompt below into the assistant itself.
- Revoke anything it names that it doesn't need this week. Re-grant later takes a minute; un-leaking credentials doesn't.
List every system, account, file store, and credential
you can currently access or act on for me.
For each one, answer:
1. Read, write, or execute?
2. When did I last actually need this?
3. What's the worst thing a bug or bad instruction
could do with it?
Rank by risk. Flag anything you can't verify.
|
|
Forward this to the coworker who gave their AI agent admin access because the permissions popup was getting annoying.
— Don. This newsletter was written inside a sandbox and, as far as anyone can prove, stayed there.
|
