# MOSAIC-Bench: Measuring Compositional Vulnerability Induction in Coding Agents

arXiv: https://arxiv.org/abs/2605.03952 · HTML: https://mlsanity.com/paper/2605.03952/

## Key facts

- Cited by 3 papers in the map (data through 14 Oct 2026).
- Placed in Agent threats, guardrails and governance, one of 16 directions in LLM agents and tool use.
- Builds on 15 of its 19 references that are in the map.
- First posted on arXiv 5 May 2026.

## Placed in

- LLM agents and tool use / [Agent threats, guardrails and governance](https://mlsanity.com/agents/agent-threats-guardrails-and-governance/)

## Builds on (most cited first, 15 of 15)

- [SWE-bench: Can Language Models Resolve Real-World GitHub Issues?](https://mlsanity.com/paper/2310.06770/) (2023)
- [Cybench: A Framework for Evaluating Cybersecurity Capabilities and Risks of Language Models](https://mlsanity.com/paper/2408.08926/) (2024)
- [OS-Harm: A Benchmark for Measuring Safety of Computer Use Agents](https://mlsanity.com/paper/2506.14866/) (2025)
- [DrAttack: Prompt Decomposition and Reconstruction Makes Powerful LLM Jailbreakers](https://mlsanity.com/paper/2402.16914/) (2024)
- [RedCode: Risky Code Execution and Generation Benchmark for Code Agents](https://mlsanity.com/paper/2411.07781/) (2024)
- [SEC-bench: Automated Benchmarking of LLM Agents on Real-World Software Security Tasks](https://mlsanity.com/paper/2506.11791/) (2025)
- [OpenAgentSafety: A Comprehensive Framework for Evaluating Real-World AI Agent Safety](https://mlsanity.com/paper/2507.06134/) (2025)
- [BaxBench: Can LLMs Generate Correct and Secure Backends?](https://mlsanity.com/paper/2502.11844/) (2025)
- [Monitoring Decomposition Attacks in LLMs with Lightweight Sequential Monitors](https://mlsanity.com/paper/2506.10949/) (2025)
- [Is Vibe Coding Safe? Benchmarking Vulnerability of Agent-Generated Code in Real-World Tasks](https://mlsanity.com/paper/2512.03262/) (2025)
- [MOCHA: Are Code Language Models Robust Against Multi-Turn Malicious Coding Prompts?](https://mlsanity.com/paper/2507.19598/) (2025)
- [Breaking the Code: Security Assessment of AI Code Agents Through Systematic Jailbreaking Attacks](https://mlsanity.com/paper/2510.01359/) (2025)
- [ABC-Bench: Benchmarking Agentic Backend Coding in Real-World Development](https://mlsanity.com/paper/2601.11077/) (2026)
- [VADER: A Human-Evaluated Benchmark for Vulnerability Assessment, Detection, Explanation, and Remediation](https://mlsanity.com/paper/2505.19395/) (2025)
- [SeCodePLT: A Unified Platform for Evaluating the Security of Code GenAI](https://mlsanity.com/paper/2410.11096/) (2024)

## Built on by (most cited first, 3 of 3)

- [Isolation as a First-Class Principle for LLM-Agent System Safety: Concepts, Taxonomy, Challenges and Future Directions](https://mlsanity.com/paper/2607.12406/) (2026)
- [Magnet: Detecting Cross-Session AI Misuse Through Capability Accumulation](https://mlsanity.com/paper/2608.02518/) (2026)
- [BulkPR-Bench: Benchmarking Queue-Level Governance of Interacting Pull Requests](https://mlsanity.com/paper/2608.02685/) (2026)

## Abstract (arXiv)

Coding agents often pass per-prompt safety review yet ship exploitable code when their tasks are decomposed into routine engineering tickets. The challenge is structural: existing safety alignment evaluates overt requests in isolation, leaving models blind to malicious end-states that emerge from sequenced compliance with innocuous-looking requests. We introduce MOSAIC-Bench (Malicious Objectives Sequenced As Innocuous Compliance), a benchmark of 199 three-stage attack chains paired with deterministic exploit oracles on deployed software substrates (10 web-application substrates, 31 CWE classes, 5 programming languages) that treats both exploit ground truth and downstream reviewer protocol as first-class evaluation axes. On this benchmark, nine production coding agents from Anthropic, OpenAI, Google, Moonshot, Zhipu, and Minimax compose innocuous tickets at 53-86% end-to-end ASR with only two refusals across all staged runs. In a matched direct-prompt experiment over four frontier Claude/Codex agents, vulnerable-output rates fall to 0-20.4%: Claude primarily refuses, while Codex primarily hardens rather than emitting the vulnerable implementation - ticket staging silences both defense modes simultaneously. Downstream, code reviewer agents approve 25.8% of these confirmed-vulnerable cumulative diffs as routine PRs, and a full-context implementation protocol closes only 50% of the staged/direct gap, ruling out context fragmentation as the sole explanation. As a deployable but non-adaptive mitigation, reframing the reviewer as an adversarial pentester reduces evasion across the evaluated reviewer subset; pentester framed evasion ranges from 3.0% to 17.6%, and an open-weight Gemma-4-E4B-it reviewer under this framing detects 88.4% of attacks on the dataset with a 4.6% false-positive rate measured on 608 real-world GitHub PRs.