The Encrypted CoT Block Had One Global Key. Every Session Shared It.
Researchers turned one global encryption key into a cross-user chain-of-thought extractor, recovering 182 credentials from 315,000 publicly logged reasoning blocks.
Benchmark Matrix
Source: Artificial Analysis · Updated 4 hours ago · Methodology
Scores are Artificial Analysis's benchmark results in their native scales (0–1 accuracy for the six axes; the composite is on AA's own internal scale). AA periodically re-baselines its benchmark set.
| # | Model | GPQA 0–1.0 | HLE 0–1.0 | SciCode 0–1.0 | AA-LCR 0–1.0 | τ³-Bank 0–1.0 | TermBench2 0–1.0 | Intelligence Index context |
|---|---|---|---|---|---|---|---|---|
| 01 | Claude Opus 5 (Adaptive Reasoning, Max Effort) Anthropic | 0.932 | 0.549 | 0.557 | 0.757 | 0.421 | 0.891 | 63.1 |
| 02 | Claude Fable 5 (Adaptive Reasoning, Max Effort, Opus 4.8 Fallback) Anthropic | 0.926 | 0.555 | 0.602 | 0.767 | 0.381 | 0.846 | 62.1 |
| 03 | GPT-5.6 Sol (max) OpenAI | 0.941 | 0.495 | 0.561 | 0.777 | 0.443 | 0.880 | 60.9 |
| 04 | Grok 4.6 (high) SpaceXAI | 0.950 | 0.429 | 0.536 | 0.750 | 0.507 | 0.884 | 60.9 |
| 05 | Kimi K3 (max) Kimi | 0.935 | 0.469 | 0.587 | 0.827 | 0.460 | 0.850 | 59.7 |
| 06 | Qwen3.8 Max Alibaba | 0.927 | 0.431 | 0.529 | 0.743 | 0.513 | 0.813 | 58.1 |
| 07 | Muse Spark 1.2 (xhigh) Meta | 0.904 | 0.455 | 0.564 | 0.833 | 0.349 | 0.802 | 56.8 |
| 08 | GPT-5.6 Terra (max) OpenAI | 0.925 | 0.429 | 0.539 | 0.797 | 0.402 | 0.880 | 56.6 |
Bars are scaled to a fixed per-column reference (0–1 for accuracy benchmarks), not to the strongest model in the table. The Intelligence Index column shows AA's overall score for context only — it is partly derived from the axes to its left, so it is not bar-rendered.
What's Moving Today
- 01 Regulation as the forcing function: Anthropic frames this as EU AI Act compliance, not a proactive safety measure — a distinction that matters for how durable or globally consistent the feature will be.
- 02 The gap between "day one" and "day 47": new models get watermarking at launch, but the timeline for retrofitting older, already-deployed models remains unspecified.
- 03 Persistence versus evasion: a text watermark designed to survive copy-paste and light editing raises the immediate question of how it holds up against paraphrasing, translation, or adversarial stripping.
- 04 Provenance as infrastructure: C2PA-signed image metadata plus invisible text marks push toward a de facto labeling standard that other vendors may now feel pressure to match.
Today's Briefing
Daily Signal View archive →
Anthropic to Embed Watermarks in Claude Text and Images
Anthropic announced that Claude will embed imperceptible watermarks in generated text and attach signed provenance metadata to generated images, a move tied directly to EU AI Act transparency rules that took effect August 2. New Claude models will carry the marks from…
Read briefing →Foundational analysis across the forces shaping AI
Explore all →AI Infrastructure
Compute, chips, memory — the physical layer beneath every model.
37 articlesModel Development
Frontier research, capabilities, and how models actually improve.
73 articlesAgents & Automation
Autonomous systems, tooling, and the workflows they reshape.
34 articlesPolicy & Regulation
Governance, law, and the rules forming around AI.
44 articlesBusiness & Markets
Capital, companies, and the economics of the buildout.
86 articlesLatent Organic
Our long game — collective reasoning for AI alignment.
2 articles