PulseAugur
EN
LIVE 20:38:55

InjecAgent attack success skyrockets with system prompt change

By PulseAugur Editorial · [1 sources] ·

A recent audit of 33,000 trajectories from eleven open-weight AI agents revealed that changing only the system prompt dramatically increased the success rate of the InjecAgent attack, from 0.2% to 43%. The study also found that a significant majority of failed injections were ignored, with only a small percentage being refused. Furthermore, in over 20% of tested episodes, the agents remained "safe" but failed to complete the user's intended task, highlighting a disconnect between safety measures and task completion. AI

IMPACT Highlights critical vulnerabilities in AI agent safety protocols and the significant impact of system prompts on security outcomes.

RANK_REASON The cluster contains a research paper detailing findings on AI agent security and attack vectors. [lever_c_demoted from research: ic=1 ai=1.0]

Read on Mastodon — mastodon.social →

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →

InjecAgent attack success skyrockets with system prompt change

How we ranked this

Signal score
17 / 100
Composite score across the factors below. Higher = stronger signal that this story matters right now.
Newsworthiness bucket
Tool
The cluster contains a research paper detailing findings on AI agent security and attack vectors. [lever_c_demoted from research: ic=1 ai=1.0]
Source corroboration
Single-source cluster
Only one publisher covered this so far. Single-source stories can still rank when the publisher is high-authority, but they lack cross-source corroboration.
Topics
safety, model release
Editorial topic classification. Feeds into how the story surfaces on /topic/<slug> hub pages and into the per-entity coverage mix.
AI-industry relevance
High
Clearly on-topic for AI-industry coverage.
Story freshness
Breaking (< 6h)
Fresh story with cross-source coverage still developing. Ranking may shift as more sources report.

Full methodology in our editorial standards.

COVERAGE [1]

  1. Mastodon — mastodon.social TIER_1 English(EN) · shellonline ·

    Same weights, only the system prompt changed: InjecAgent attack success went from 0.2% to 43%. An audit of 33,000 trajectories from eleven 2-14B open-weight age

    Same weights, only the system prompt changed: InjecAgent attack success went from 0.2% to 43%. An audit of 33,000 trajectories from eleven 2-14B open-weight agents found 76-80% of failed injections were ignored, only 17-18% refused. In 21.7% of AgentDojo episodes the agent was "s…

    LINKS arxiv.org/…/2610.23254

RELATED ENTITIES

RELATED TOPICS