← The Wire

Research2026-06-27 · source-backed

Frontier model "manipulation" is task-dependent, not a single trait.

Confidence
source-backed
Sources
1
Redaction
passed

Story

A June 25 paper (arXiv:2606.25899) argues manipulation capability varies sharply by task rather than being one measurable scalar. (arXiv) That complicates any safety eval trying to score persuasion as a global number. For anyone deploying agents, the practical implication is that red-teaming has to be task-specific. A model that's harmless on one workload can be dangerous on another, and a single safety rating hides that.


Related stories

Each link below shares sources, entities, or timing with this story.

  1. Shared entity: Frontier / Shared topic / What happened next

    OpenAI is sunsetting Agent Builder and Evals, with Frontier as the go-forward platform.

    Both cover Frontier; overlapping topics (agent, deploying, eval, frontier); picks up the Frontier thread on 2026-06-28.

  2. Shared entity: Frontier / Shared topic / Earlier coverage / Tension

    OpenAI Frontier: Enterprise Agent Platform Directly Competes With Incumbents

    Both cover Frontier; overlapping topics (agent, deploying, frontier); earlier Frontier coverage from 2026-02-25.

  3. Shared entity: Frontier / Same source domain / Shared topic / Earlier coverage

    Benchmarks & Evaluation

    Both cover Frontier; reported by the same outlet (arxiv.org); overlapping topics (agent, frontier).

  4. Shared entity: Frontier / Shared topic / What happened next

    Frontier capability and general availability are now decoupled by policy.

    Both cover Frontier; overlapping topics (capability, frontier, model); picks up the Frontier thread on 2026-06-29.

  5. Shared entity: Frontier / Shared topic / Earlier coverage / Tension

    GPT-5.5 Is Here. It's Fast, It's Expensive, and It Changes How You Prompt.

    Both cover Frontier; overlapping topics (capability, model); earlier Frontier coverage from 2026-04-25.

  6. Shared entity: Frontier / Shared topic / Earlier coverage

    Devstral 2 is open source, 123B params, hits 72.2% SWE-bench, and costs a seventh of Sonnet

    Both cover Frontier; overlapping topics (agent, frontier, model); earlier Frontier coverage from 2026-06-02.

  7. Cursor Proves the Commoditization Thesis: Composer 2.5 Matches Opus 4.7 at $0.50 Per Million Tokens

    Both cover Frontier; overlapping topics (agent, frontier, model); earlier Frontier coverage from 2026-05-25.

  8. Same source domain / Shared topic / Tension

    Top coding agents secretly write Python to write your obscure language for you.

    Reported by the same outlet (arxiv.org); overlapping topics (agent, capability, model); pushes against this story (but).

Source trail

Entities

Provenance

AI generated
no
Story unit
2026-06-27-frontier-model-manipulation-is-task-dependent-not-a-single-trait
Labels
source-backed, canonical briefing excerpt