The worst agent attacks now teach themselves. a september 3 study shows the sir attack lifting…
the worst agent attacks now teach themselves. a september 3 study shows the sir attack lifting hijack success on computer-use agents from 0% to 28% on gemini 3.5 flash and 4% to 24% on claude opus 4.8, learning by trial and error while the agent still finishes its task.
the exploit improves faster than the guardrail.
Context
SIR: Self-improving Red-teaming for Computer Use Agents (arXiv 2608.30207) describes a black-box attacker that plants injections in an untrusted region such as a forum post, cannot change the user instruction or the OS, and learns from failed trajectories using compositional search and failure-driven strategy discovery. It tests 50 sampled cases from RedTeamCUA on three agents (Claude Opus 4.6, Opus 4.8 and Gemini 3.5 Flash). Hijack success is joint success: the adversarial objective and the benign task are both completed.
Gemini 3.5 Flash from 0 to 28 percent matches the abstract and body. For Opus 4.8 the abstract and contribution list say 4 to 22 percent while the results section says 4 to 24 percent; the paper is inconsistent and the note's 24 matches one place only. Still finishes its task matches the joint-success definition. A September 3 study was not dated in the text read; the arXiv ID and a snippet date of 31 August 2026 point to late August, so the date is unverified. Small sample of 50 cases on one benchmark, and not an independent replication.
Watch next
- Replication on other computer use agents.
Sources
Provenance
The note above is reproduced unedited from the original post, first published on Threads on 20 September 2026 at 20:48 IST. Sources are the papers and datasets the note draws on.
View the original post ↗Embed this note
More notes
The air is now being asked to keep its own ledger
the air is now being asked to keep its own ledger: ecmwf’s aifs compo becomes the first ai model to forecast atmospheric composition globally every three hours, cleanair simulates 365 days of pm2.5 over china in ten seconds, and a unified framework maps six pollutants at one kilometer across the whole country. the air now files its own composition report.
read the note →The current is now being asked to draw its own map
the current is now being asked to draw its own map: china’s langya 2.0 predicts six ocean phenomena including internal waves and mesoscale eddies, a deep net called wenhai resolves eddies globally with air sea flux formulas built in, and scripps infers surface currents from the way temperature patterns deform in satellite images. the ocean now files its own circulation report.
read the note →The soil is now being asked to report its own carbon
the soil is now being asked to report its own carbon: a nix color sensor paired with generative data augmentation predicts soil organic carbon without a lab, random forest drives 74 percent of soil health mapping studies, and sentinel 2 tracks five year carbon change across france and italy from 922 samples. the dirt now files its own carbon account.
read the note →