Linear probes fail to detect prompt injection without matched negative training
Source published ·Modelwire updated
Original coverage: arXiv cs.CL ↗·How Modelwire adds context
The development
Researchers expose a critical flaw in how linear probes detect prompt injection attacks on LLM agents. Current evaluation methods conflate surface-level patterns like conversation position and text length with actual injection detection, achieving high scores without understanding the attack itself. By introducing matched negatives into training data, the team demonstrates that standard probes fail to distinguish injected content from harmless text formatted similarly. This work matters for production deployments relying on cheap monitoring layers to flag suspicious tool outputs before human review, revealing that many false positives stem from probe overfitting rather than genuine security capability.
Modelwire’s AI-generated summary of coverage from arXiv cs.CL.
Modelwire analysis
Skeptical readOur AI-generated reading of the wider context and the next developments to watch.
The paper doesn't just show that probes fail at injection detection. It reveals that high benchmark scores on current datasets are largely artifacts of overfitting to conversation structure rather than evidence of actual security capability. This means existing production deployments may be reporting false confidence.
This connects directly to the TRACE work from October 8th, which exposed how verifier brittleness can mask evaluation artifacts as real capability differences. Here, linear probes achieve high scores not because they detect injections but because they've learned to exploit superficial patterns like position and length. The same mechanistic blindness appears across both agent evaluation and injection detection. Both papers suggest that current benchmarking practices systematically conflate measurement artifacts with genuine security or performance signals.
If the researchers release a standardized injection detection benchmark that controls for conversation position and text length, and existing commercial probe vendors' scores drop by more than 15 percentage points on that benchmark, it confirms this was a real gap in production monitoring rather than a lab-specific artifact.
This interpretation is generated from the summary above and the archive coverage cited below. Our methodology · Report an error
Coverage behind this analysis
These archive entries ground the connection in our analysis. They are ordered by source publication date, with links to our coverage and the original sources.
·arXiv cs.CL
Verifier scores mask evaluation brittleness in LLM agent benchmarks
Researchers have exposed a critical vulnerability in how LLM agent performance is measured: verifier scores can shift dramatically based on superficial evaluation changes rather than actual capability differences. TRACE, a diagnostic protocol, systematically isolates whether score movements reflect genuine agent behavior or artifacts of the scoring mechanism itself. In controlled experiments, renaming tools alone caused…
MentionsLinear probes · Prompt injection · LLM agents · Tool outputs
How this coverage is produced
Modelwire uses AI to generate summaries and context from source headlines, snippets, and selected archive coverage. Automated checks do not verify every claim, and items are not routinely reviewed by a person before publication. Zacaria Solis operates the site. Read the linked source for the full evidence and report errors through our corrections process.
Modelwire summarizes, we don’t republish. arXiv cs.CL originally reported this story as “Prompt-Injection Probes Learn the Conversation, Not Just the Injection”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.