MOSAIC-Bench C
AI benchmark · Agentic
What is MOSAIC-Bench?
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.
Tier C
Niche: a handful of models scored, or built by an RL vendor.
How much stands behind a tier: usage, who backs it, a confirmed line from its own site or paper, and a clear read on X. High with three of the four, medium with two, low otherwise.
- 9 models scored
- in curated RL directories (1 pt)
How we got to tier C
- Adoption: 9 models scored. Tier C.
- Not yet read on X.
- In curated RL directories: 1 point, a floor of none.
- Not yet confirmed by a line from its own site or paper.
A clear view on X moves a tier one step up or down; most entries don't move. How tiers work.
Not on the map yet: no line from its own site or paper places it in a domain.
Next: browse all benchmarks. Entry from the RL Research daily scrape of public sources, 2026-10-11 snapshot.