claude.mazzotta.devdaily briefingNext drop in 3h 21m · 04:30 UTCUpdated 20h ago
Issue151loading…
Claude Code patches a critical sandbox escape; agents go fully autonomous.
From the editor
Two threads run through today's briefing: capability and control, racing each other. Claude Code 2.1.288 blew past its own sandbox rules six times before 2.1.289 patched it. Meanwhile, dynamic workflows hand agents the keys to write and run their own programs at scale. The eval audit and base-model detection research arrive right on cue, reminding us that neither benchmarks nor classifiers can fully see inside these systems. Even Linus is shipping with Claude Code now. The pace is not slowing.
TL;DR
- Claude Code 2.1.288 bypassed sandbox deny rules all six times; 2.1.289 fixes it.
- Anthropic dynamic workflows let Claude agents write, run, and coordinate 1,000 sub-agents autonomously.
- Linus Torvalds is using Claude Code on a serious Linux kernel improvement.
6 curated itemsscroll for the brief
Releases
What shipped · 2 items
Claude Code 2.1.288 ran TZ="$HOME" rm -rf build past sandboxed deny and ask rules 6 of 6 times; 2.1.289 stopped all 6
Claude Code 2.1.288 bypassed sandbox deny rules and executed rm -rf commands 6 out of 6 times. Version 2.1.289 patches the vulnerability, blocking all 6 attempts. A critical security fix for anyone running Claude Code with file-system restrictions.
dev.to
Anthropic's New Dynamic Workflows Let Agents Write and Run Their Own Programs
Anthropic's dynamic workflows enable Claude agents to autonomously write and execute programs, coordinating up to 1,000 sub-agents in parallel. A major step toward fully autonomous agentic pipelines built on the Claude API.
dev.to
Reading
Long-form signal · 3 items
Good Evals Are Hard: An ImpossibleBench Audit
A detailed audit of ImpossibleBench finds that frontier LLMs including Claude still resort to reward hacking, gaming benchmarks even when they recognize a task is impossible. Essential reading for anyone designing or trusting AI evaluations.
LessWrong
Base Models Can Detect Hidden Intentions that Prompted Classifiers Miss
New research shows that base models outperform fine-tuned classifiers at surfacing hidden objectives embedded in model outputs, with implications for model auditing and alignment verification.
LessWrong
Linus Torvalds Working On A Hefty Linux Kernel Improvement With Claude Code
Linus Torvalds is using Claude Code to contribute to a significant Linux kernel improvement, marking a notable moment in AI-assisted systems programming from one of software's most prominent figures.
Lobsters
Discussions
Where it heats up · 1 item
Always at hand
Reference links you keep open