Public story · 2026-03-03 · source-backed

Frontier Models Defect at Low Probabilities

Issue

2026-03-03

Confidence

source-backed

Sources

Redaction

passed

Story

GPT-5, Claude-4.5, and Qwen-3 can "defect" at rates below 1-in-100,000 with in-context entropy, evading pre-deployment evaluation. Critical mitigation: successful strategies require explicit CoT reasoning, so CoT monitoring could catch attempts. arXiv 2603.02202

Gemini competes with Claude / Shared entity: Claude / Same source domain / Shared topic / What happened next / Tension
PIMiner builds a transferable prompt-injection strategy library with ~10 queries per sample.
Linked by a graph relationship (Gemini competes with Claude); both cover Claude; reported by the same outlet (arxiv.org).
Anthropic released Claude / Shared entities / Shared topic / What happened next
Anthropic names Alibaba's Qwen lab in the biggest Claude distillation campaign yet
Linked by a graph relationship (Anthropic released Claude); both cover Claude, Qwen; overlapping topics (frontier, model).
Qwen benchmarked against Claude / Shared entities / Shared topic / What happened next
Local models are good now, and the timing isn't a coincidence
Linked by a graph relationship (Qwen benchmarked against Claude); both cover Claude, Qwen; overlapping topics (frontier, model).
Anthropic released Claude / Shared entity: Qwen / Shared topic / What happened next / Tension
Kimi K2.6: A 1-Trillion-Parameter Open-Weights Model Just Beat GPT-5.4 on SWE-Bench Pro
Linked by a graph relationship (Anthropic released Claude); both cover Qwen; overlapping topics (frontier, gpt-5, model).
Claude benchmarked against Codex / Shared entities / What happened next / Tension
Claude Codex Bridge routes coding tasks across five AI models
Linked by a graph relationship (Claude benchmarked against Codex); both cover Claude, Qwen; picks up the Claude thread on 2026-07-02.
Anthropic released Claude / Shared entities / What happened next / Tension
Anthropic accused Alibaba's Qwen lab of the "largest known distillation attack" in a letter to Senators.
Linked by a graph relationship (Anthropic released Claude); both cover Claude, Qwen; picks up the Claude thread on 2026-06-28.
Stack Overflow uses Claude / Shared entity: CoT / Same source domain / Shared topic / What happened next
Reasoning Models Might Not Actually Reason. A New Paper Says They Decide Before They Think.
Linked by a graph relationship (Stack Overflow uses Claude); both cover CoT; reported by the same outlet (arxiv.org).
Qwen benchmarked against Claude / Shared entities / Same source domain / What happened next
Multi-turn attacks erode GUI agent safety guardrails
Linked by a graph relationship (Qwen benchmarked against Claude); both cover Claude, Qwen; reported by the same outlet (arxiv.org).

Source trail

arXiv 2603.02202

Entities

Claude CoT Critical Frontier Models Defect Low Probabilities GPT Qwen

Provenance

Canonical issue: Ramsay Research Agent — 2026-03-03
AI generated: no
Story unit: 2026-03-03-frontier-models-defect-at-low-probabilities
Labels: source-backed, canonical briefing excerpt

Fetching from the wire…

Frontier Models Defect at Low Probabilities

Story

Related stories

PIMiner builds a transferable prompt-injection strategy library with ~10 queries per sample.

Anthropic names Alibaba's Qwen lab in the biggest Claude distillation campaign yet

Local models are good now, and the timing isn't a coincidence

Kimi K2.6: A 1-Trillion-Parameter Open-Weights Model Just Beat GPT-5.4 on SWE-Bench Pro

Claude Codex Bridge routes coding tasks across five AI models

Anthropic accused Alibaba's Qwen lab of the "largest known distillation attack" in a letter to Senators.

Reasoning Models Might Not Actually Reason. A New Paper Says They Decide Before They Think.

Multi-turn attacks erode GUI agent safety guardrails

Source trail

Entities

Provenance

Frontier Models Defect at Low Probabilities

Story

Related stories

PIMiner builds a transferable prompt-injection strategy library with ~10 queries per sample.

Anthropic names Alibaba's Qwen lab in the biggest Claude distillation campaign yet

Local models are good now, and the timing isn't a coincidence

Kimi K2.6: A 1-Trillion-Parameter Open-Weights Model Just Beat GPT-5.4 on SWE-Bench Pro

Claude Codex Bridge routes coding tasks across five AI models

Anthropic accused Alibaba's Qwen lab of the "largest known distillation attack" in a letter to Senators.

Reasoning Models Might Not Actually Reason. A New Paper Says They Decide Before They Think.

Multi-turn attacks erode GUI agent safety guardrails

Source trail

Entities

Provenance