Fetching from the wire…
Public story · 2026-07-27 · high
Whitelist controls and LLM review narrowed the damage but neither stopped the injection outright, the paper found.
Why now: Covered in the July 27 briefing, with no separate publication date given.
A compromised router beat every coding agent tested, posting a 0% defense rate across four escalating injection levels, per arXiv 2607.23624.
Third-party routers like OpenRouter sit on the trusted path between an agent and its model. Nothing on that path verifies the response reaching the agent matches what the provider sent, and the agent executes on it anyway.
The researchers built SIDEL, a trace record-replay-inject framework. They tested it against four coding agents, using a 400-sample dataset across four escalating injection levels.
Two mitigations, whitelist-based execution control and LLM-based review, cut into how much damage got done, but the defense success rate held at 0%. Neither restored end-to-end control once the router had tampered with the response.
I've recommended routing through gateways like OpenRouter. This paper is the bill for that advice landing in my own feed.
What I'd want to see next is router-side attestation, something that proves the response reaching the agent is the one the provider actually sent. The framework's code is posted under Riyasushin's SIDE repo on GitHub.
Each link below shares sources, entities, or timing with this story.
Simon Willison found it in the OpenRouter price list. 1.6T MoE, ~49B active, 1M context, up to 384K output, at $0.435/M input (cache miss) and $0.87/M output. The agentic-coding deltas versus the preview are the story: DeepSWE 12.8 → 62.7, CyberGym 52.7 → 83.3, Terminal Bench...
1. Set package cooldown to 72 hours across all your package managers. pnpm: resolution-time=72h, uv: --exclude-newer, npm via .npmrc. This single config change would have protected you from the LiteLLM attack. Willison's survey covers all seven managers. 2. Install Lasso Secur...
Allen Bargi's August 15 post hit 302 points arguing that AI collaboration rewards context-sharing, examples, and feedback over precise instruction (Hacker News). The pushback holds that the piece conflates management with leadership. mikeocool calls it "the most low effort ver...
The WSJ reported July 24 that Stripe is negotiating for the model-routing marketplace at roughly $10 billion against the ~$1.3 billion OpenRouter carried after a May round. OpenRouter already runs its own billing on Stripe. This would be Stripe's largest move beyond payments,...
Stripe published Part 2 of its Minions engineering blog, and it's the most detailed production agent architecture I've read from any company this year. The numbers alone are worth the read: 1,300+ weekly merged PRs from coding agents. But the architecture decisions matter more...
— "The defining characteristic of a coding agent is that it can execute the code it writes." Never assume LLM-generated code works without verification. Patterns for python -c edge case testing, /tmp demo files, browser automation with Playwright/Rodney. Red/green TDD: when ag...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.