threatdeck
Live 14:45 UTC ·

New Technique Medium TD-2026-0086 For AI Eng, AppSec, Red Team, CISO AI-assisted

S1-Evo: self-evolving attacks manipulate typed System-1 decision model outputs

The paper proposes S1-Evo, a framework that perturbs untrusted fields in inputs to typed System-1 decision models so that the models' typed outputs and execution branches change. Across four benchmarks and three S1 models, the authors report 71.99% mean attack success rate and 57.47% high-confidence ASR on jev-1.14, and up to 24.00% end-to-end agent violations.

Why it matters

LLM agents that use typed S1 models for routing, moderation, or authorization can have those decisions altered by untrusted context. An attacker who controls such context could change which execution branch an agent takes.

Do this

  1. Inventory LLM agent components that use typed classifier or scoring models for routing, moderation, or authorization decisions.
  2. Treat untrusted fields fed to S1 decision models as adversarial input and validate or sanitize them before scoring.
  3. Add independent checks on high-impact authorization or routing decisions instead of relying on a single S1 model output.
  4. Log S1 option scores and margins so that sudden near-tie or flipped decisions can be reviewed.
  5. Re-evaluate S1 models against contract-aware perturbation tests once the authors release their evaluation code.

Signals

  • NoCISA KEV
  • 23/100Priority
  • 60%Confidence

Evidence

“Across four benchmarks and three S1 models, S1-Evo reaches 71.99\% mean ASR and 57.47\% high-confidence ASR on \texttt{jev-1.14}, and causes up to 24.00\% E2E agent violations.”

arXiv (AI security papers)

“yet untrusted context can still manipulate typed outputs and alter execution branches”

arXiv (AI security papers)

Sources

Affected / related

S1-Evo, jev-1.14

Published 2026-11-12 15:21 UTC · updated 2026-11-13 07:05 UTC · AI-assisted summary, checked against quoted sources · How we score

Following

Follow the vendors, products, AI tools, techniques and actors you care about. Matching stories, CVEs and exploited bugs are marked, a Following tab appears in the feed, and returning visits show what changed. Saved only in this browser.

Presets

Get updates

A personal RSS feed of new stories on everything you follow. Paste it into any feed reader, Slack or Teams RSS app.