Fetching from the wire…
Agents2026-07-18 · source-backed
arXiv 2607.02514 finds ≥65% evasion generalizes across attacker backends (Sonnet 4.5, Gemini 3.1 Pro, Kimi K2.5) and across state-of-the-art monitor models. The win: a four-monitor ensemble combining a stateful link-tracker with trajectory monitors drops gradual-attack evasion from 93% (weakest standard diff monitor) to 47%. Diff-level review of each step is not enough when state persists between steps. 47% is still terrible, to be clear. It's just half as terrible.
Each link below shares sources, entities, or timing with this story.
Claude Code uses Sonnet / Shared entities / Earlier coverage / Tension
Linked by a graph relationship (Claude Code uses Sonnet); both cover Gemini, Sonnet; earlier Gemini coverage from 2026-05-17.
Gemini built by Google / Shared entities / Earlier coverage
Linked by a graph relationship (Gemini built by Google); both cover Gemini, Kimi K2; earlier Gemini coverage from 2026-06-12.
Claude Code uses Sonnet / Shared entities / Earlier coverage
Linked by a graph relationship (Claude Code uses Sonnet); both cover Gemini, Sonnet; earlier Gemini coverage from 2026-03-28.
Linked by a graph relationship (Claude Code uses Sonnet); both cover Gemini, Kimi K2; earlier Gemini coverage from 2026-02-24.
Linked by a graph relationship (Claude Code uses Sonnet); both cover Diff, Gemini; earlier Diff coverage from 2026-02-25.
Copilot deprecates Gemini / Shared entity: Gemini / Same source domain / Earlier coverage
Linked by a graph relationship (Copilot deprecates Gemini); both cover Gemini; reported by the same outlet (arxiv.org).
Claude Code uses Sonnet / Shared entity: Gemini / Same source domain / Earlier coverage
Linked by a graph relationship (Claude Code uses Sonnet); both cover Gemini; reported by the same outlet (arxiv.org).
Gemini competes with ChatGPT / Shared entities / Earlier coverage
Linked by a graph relationship (Gemini competes with ChatGPT); both cover Gemini, Kimi K2; earlier Gemini coverage from 2026-06-21.