Instruction Bleed: Cross-Module Interference in Prompt-Composed Agentic Systems
Quick Answer
The study identifies compositional behavioral leakage (CBL) in prompt-composed systems, where editing one module affects others without direct dependencies.
Quick Take
Testing on Claude Sonnet 4.6 revealed significant interference through content changes, highlighting the need for cross-module interference measurement in .
Key Points
- CBL occurs due to architectural non-isolation in transformer models.
- Testing involved 144 trials with Claude Sonnet 4.6 using a three-channel protocol.
- Only content changes produced a detectable effect (Cohen's d = 0.63).
- CBL is distinct from known agent-failure issues like adversarial injection.
- The study provides a reusable protocol and operational definitions for future research.
Paper Resources
📖 Reader Mode
~2 min readAbstract:Practitioners of prompt-composed agentic systems report a recurring failure mode: editing one prompt module silently shifts the behavior of others despite no shared variable or executable dependency. We formalize this as compositional behavioral leakage (CBL): interference between modules sharing a context window. CBL is enabled by architectural non-isolation: transformer self-attention provides no formal boundary between concatenated modules. We probe CBL on a deployed job-evaluation agent (Claude Sonnet 4.6, 144 trials) through a reusable three-channel protocol that perturbs non-focal modules along volume, content, and form. Only the content channel produces a detectable paired effect (Cohen's d = 0.63, bootstrap 95% CI excluding zero); no recommendation flipped -- a sub-threshold regime invisible to standard QA but compounding across the thousands of decisions a deployed agent makes. CBL is orthogonal to known agent-failure axes (adversarial injection, cognitive degradation, multi-agent fault propagation, privacy leakage). We contribute an operational definition, a reusable protocol, a falsifiable prediction set, and a system-class characterization, establishing cross-module interference measurement as a requirement for prompt-composed agent evaluation.
| Comments: | 8 pages, 2 tables. Accepted to the ICML 2026 Workshop on Failure Modes in Agentic AI (FAGEN), Seoul, South Korea |
| Subjects: | Artificial Intelligence (cs.AI); Information Retrieval (cs.IR); Multiagent Systems (cs.MA) |
| Cite as: | arXiv:2606.26356 [cs.AI] |
| (or arXiv:2606.26356v1 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2606.26356 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Ching-Yu Lin [view email]
[v1]
Wed, 24 Jun 2026 20:09:28 UTC (39 KB)
— Originally published at arxiv.org
Want this in your inbox every morning?
Daily brief at your local 8am — bilingual EN/中文, free.
More from arXiv cs.AI
See more →HOBA: Hierarchical On-Policy Bidding Agents for Adaptive Online Advertising
HOBA (Hierarchical On-policy Bidding Agents) is a novel hierarchical reinforcement learning framework that enhances online advertising bidding systems by improving adaptability and reducing hyperparameter tuning costs. It utilizes a for hyperparameter inference, a SARSA agent for expert model selection, and a dynamic expert pool for bid execution, achieving a +3.6% increase in target cost during large-scale deployment and outperforming state-of-the-art baselines on AuctionNet.