Augur Dispatch

Chain of evidence

Evidence for 2026-08-28

This frozen page shows Augur's claims and source links for one sent dispatch. Stored spot-checks appear only where the frozen edition supports them; absence is not presented as verification.

As of:

Bundle identity: evidence-bundle-v1-1b88c2789ba112fc3760265af76eb4706e0475d035b9cb404ef2c8285f2ddbc2

Format: evidence-bundle-v1 · 26 claims

Assertion 1

Its auto mode replaces the constant "may I run this?" prompts with model-based classifiers, small AI judges that approve or block each action, and Anthropic turned that on by default for paid plans on August 14 Anthropic Engineering Latent Space.

Assertion status: No spot-check verdict is published for this assertion.

Claude Code introduces an 'auto mode' that delegates permission approvals to model-based classifiers to reduce user fatigue from manual prompts.

Claim 25661 Label: fact Provenance: primary Corroboration eligibility unknown · not recorded in this bundle Recorded

Anthropic Engineering

No stored spot-check names this claim in this edition.

Auto mode is intended to block dangerous actions caused by overeager behavior, honest mistakes, prompt injection, and model misalignment.

Claim 25666 Label: fact Provenance: primary Corroboration eligibility unknown · not recorded in this bundle Recorded

Anthropic Engineering

No stored spot-check names this claim in this edition.

Anthropic rolled out Auto mode as the default permissions mode in Claude Code for Pro/Max/Team, with repo-aware setup to suggest trusted repositories and domains.

Claim 48027 Label: fact Provenance: primary Corroboration eligibility unknown · not recorded in this bundle Recorded

Latent Space

No stored spot-check names this claim in this edition.

Assertion 2

Anthropic ran a study with 1,053 paid testers, and auto mode caught 89% of harmful actions while human review caught 13.6%, largely because annoyed users click approve without reading TechCrunch AI.

Assertion status: No spot-check verdict is published for this assertion.

Anthropic will make Claude Code's auto mode the default for Pro, Max, and Team accounts starting on August 14, 2026.

Claim 44361 Label: fact Provenance: primary Corroboration eligibility unknown · not recorded in this bundle Recorded

TechCrunch AI

No stored spot-check names this claim in this edition.

In a study of 1,053 paid testers, auto mode caught 89% of harmful actions while human review caught only 13.6%.

Claim 44362 Label: fact Provenance: primary Corroboration eligibility unknown · not recorded in this bundle Recorded

TechCrunch AI

No stored spot-check names this claim in this edition.

Anthropic claims that in testing, auto mode proved safer than manual review.

Claim 44363 Label: opinion Provenance: primary Corroboration eligibility unknown · not recorded in this bundle Recorded

TechCrunch AI

No stored spot-check names this claim in this edition.

Assertion 3

Prompt injection researcher Johann Rehberger found an attack on auto mode that he claims works 80% of the time Simon Willison's Weblog.

Assertion status: No spot-check verdict is published for this assertion.

Anthropic recently made Claude Code's auto mode the default for protecting users against prompt injection attacks.

Claim 52980 Label: fact Provenance: primary Corroboration eligibility unknown · not recorded in this bundle Recorded

Simon Willison's Weblog

No stored spot-check names this claim in this edition.

Anthropic has made bold claims about the effectiveness of Claude Code's auto mode.

Claim 52981 Label: fact Provenance: primary Corroboration eligibility unknown · not recorded in this bundle Recorded

Simon Willison's Weblog

No stored spot-check names this claim in this edition.

Prompt injection researcher Johann Rehberger found an attack against Claude Code's auto mode that he claims works 80% of the time.

Claim 52982 Label: fact Provenance: primary Corroboration eligibility unknown · not recorded in this bundle Recorded

Simon Willison's Weblog

No stored spot-check names this claim in this edition.

Assertion 4

The claim lands harder against Anthropic's own framing: Boris Cherny of Anthropic had claimed that Claude Opus 5 plus auto mode pushed injection success to near zero Don't Worry About the Vase.

Assertion status: No spot-check verdict is published for this assertion.

Boris Cherny of Anthropic claims that Claude Opus 5 is the least prompt-injectable model to date, with a near-zero success rate for prompt injection attacks when combined with strong alignment and Claude Code's Auto Mode.

Claim 38672 Label: fact Provenance: primary Corroboration eligibility unknown · not recorded in this bundle Recorded

Don't Worry About the Vase

No stored spot-check names this claim in this edition.

Assertion 5

Google DeepMind is now piloting what it calls the world's first double-blind evaluation of a proprietary frontier-class model, testing Gemini Flash Lite against confidential benchmarks with partners including the Singapore AI Safety Institute and MLCommons Google DeepMind Blog.

Assertion status: No spot-check verdict is published for this assertion.

Google DeepMind is piloting the world's first double-blind evaluation of a proprietary, frontier-class AI model.

Claim 52719 Label: fact Provenance: primary Corroboration eligibility unknown · not recorded in this bundle Recorded

Google DeepMind Blog

No stored spot-check names this claim in this edition.

The double-blind evaluation pilot involves testing the Gemini Flash Lite model against confidential benchmarks.

Claim 52720 Label: fact Provenance: primary Corroboration eligibility unknown · not recorded in this bundle Recorded

Google DeepMind Blog

No stored spot-check names this claim in this edition.

Google DeepMind is partnering with the Singapore AI Safety Institute, OpenMined, AVERI, and MLCommons for this evaluation pilot.

Claim 52721 Label: fact Provenance: primary Corroboration eligibility unknown · not recorded in this bundle Recorded

Google DeepMind Blog

No stored spot-check names this claim in this edition.

Assertion 6

OpenAI has now released a post-mortem, its account of what failed, alongside promised security and alignment fixes, and METR and Redwood Research supplied partial external analysis Don't Worry About the Vase OpenAI News.

Assertion status: No spot-check verdict is published for this assertion.

OpenAI is publishing its findings regarding the security incident involving Hugging Face.

Claim 51977 Label: fact Provenance: primary Corroboration eligibility unknown · not recorded in this bundle Recorded

OpenAI News

No stored spot-check names this claim in this edition.

OpenAI is implementing measures to enhance the security, monitoring, and alignment of its AI models.

Claim 51978 Label: fact Provenance: primary Corroboration eligibility unknown · not recorded in this bundle Recorded

OpenAI News

No stored spot-check names this claim in this edition.

OpenAI released a post-mortem report regarding an incident where its internal model hacked HuggingFace.

Claim 52776 Label: fact Provenance: primary Corroboration eligibility unknown · not recorded in this bundle Recorded

Don't Worry About the Vase

No stored spot-check names this claim in this edition.

METR and Redwood Research provided partial external analysis of the OpenAI-HuggingFace incident.

Claim 52777 Label: fact Provenance: primary Corroboration eligibility unknown · not recorded in this bundle Recorded

Don't Worry About the Vase

No stored spot-check names this claim in this edition.

Bill Gates warned of an 'economic catastrophe' related to AI.

Claim 52780 Label: fact Provenance: primary Corroboration eligibility unknown · not recorded in this bundle Recorded

Don't Worry About the Vase

No stored spot-check names this claim in this edition.

Assertion 7

Ai2 and the Providence Swedish Cancer Institute have partnered around AutoDiscovery, Ai2's tool for generating research hypotheses from data, and a joint team using it found evidence of a stronger immune signature in invasive lobular breast cancer than previously recognized Allen Institute for AI.

Assertion status: No spot-check verdict is published for this assertion.

Ai2 and the Providence Swedish Cancer Institute have formed a partnership to advance AI-assisted scientific discovery. (Source: Ai2 blog post, published 2026-08-27)

Claim 52691 Label: fact Provenance: primary Corroboration eligibility unknown · not recorded in this bundle Recorded

Allen Institute for AI

No stored spot-check names this claim in this edition.

A joint research team led by Dr. Kelly Paulson, Dr. Sasha Stanton, and Bodhisattwa Majumder used the AutoDiscovery system to analyze a breast cancer dataset. (Source: Ai2 blog post, published 2026-08-27)

Claim 52692 Label: fact Provenance: primary Corroboration eligibility unknown · not recorded in this bundle Recorded

Allen Institute for AI

No stored spot-check names this claim in this edition.

The joint research team found evidence of a stronger immune signature in invasive lobular breast cancer than previously recognized. (Source: Ai2 blog post, published 2026-08-27)

Claim 52693 Label: fact Provenance: primary Corroboration eligibility unknown · not recorded in this bundle Recorded

Allen Institute for AI

No stored spot-check names this claim in this edition.

Assertion 8

- OpenRouter's usage data shows GPT-5.6 Luna token use rose 13.8 times during OpenAI's July-to-August discount window, a demand jump worth rechecking once full prices return. OpenRouter Blog

Assertion status: No spot-check verdict is published for this assertion.

OpenAI introduced large discounts on its new Terra and Luna models from July 27th through August 14th.

Claim 52436 Label: fact Provenance: primary Corroboration eligibility unknown · not recorded in this bundle Recorded

OpenRouter Blog

No stored spot-check names this claim in this edition.

During the discount window, daily Terra token usage rose 5.6 times compared to pre-period averages.

Claim 52437 Label: fact Provenance: primary Corroboration eligibility unknown · not recorded in this bundle Recorded

OpenRouter Blog

No stored spot-check names this claim in this edition.

During the discount window, daily Luna token usage rose 13.8 times compared to pre-period averages.

Claim 52438 Label: fact Provenance: primary Corroboration eligibility unknown · not recorded in this bundle Recorded

OpenRouter Blog

No stored spot-check names this claim in this edition.

Assertion 9

- Anthropic's next Claude Code update will show whether the auto mode classifier gets retrained or the default retreats. Simon Willison's Weblog

Assertion status: No spot-check verdict is published for this assertion.

Anthropic recently made Claude Code's auto mode the default for protecting users against prompt injection attacks.

Claim 52980 Label: fact Provenance: primary Corroboration eligibility unknown · not recorded in this bundle Recorded

Simon Willison's Weblog

No stored spot-check names this claim in this edition.

Anthropic has made bold claims about the effectiveness of Claude Code's auto mode.

Claim 52981 Label: fact Provenance: primary Corroboration eligibility unknown · not recorded in this bundle Recorded

Simon Willison's Weblog

No stored spot-check names this claim in this edition.

Prompt injection researcher Johann Rehberger found an attack against Claude Code's auto mode that he claims works 80% of the time.

Claim 52982 Label: fact Provenance: primary Corroboration eligibility unknown · not recorded in this bundle Recorded

Simon Willison's Weblog

No stored spot-check names this claim in this edition.

Assertion 10

- Results from the Gemini Flash Lite confidential-benchmark pilot will show whether double-blind scoring survives contact with a real frontier model. Google DeepMind Blog

Assertion status: No spot-check verdict is published for this assertion.

Google DeepMind is piloting the world's first double-blind evaluation of a proprietary, frontier-class AI model.

Claim 52719 Label: fact Provenance: primary Corroboration eligibility unknown · not recorded in this bundle Recorded

Google DeepMind Blog

No stored spot-check names this claim in this edition.

The double-blind evaluation pilot involves testing the Gemini Flash Lite model against confidential benchmarks.

Claim 52720 Label: fact Provenance: primary Corroboration eligibility unknown · not recorded in this bundle Recorded

Google DeepMind Blog

No stored spot-check names this claim in this edition.

Google DeepMind is partnering with the Singapore AI Safety Institute, OpenMined, AVERI, and MLCommons for this evaluation pilot.

Claim 52721 Label: fact Provenance: primary Corroboration eligibility unknown · not recorded in this bundle Recorded

Google DeepMind Blog

No stored spot-check names this claim in this edition.

Assertion 11

- OpenAI's randomized study of more than 1,000 students pairs ChatGPT with critical-thinking training, and the full results deserve a close read before any school changes policy. OpenAI News

Assertion status: No spot-check verdict is published for this assertion.

OpenAI published a study examining how ChatGPT, critical thinking training, originality, and student performance on a real-world university assignment relate to each other.

Claim 52701 Label: fact Provenance: primary Corroboration eligibility unknown · not recorded in this bundle Recorded

OpenAI News

No stored spot-check names this claim in this edition.

The study described by OpenAI is a randomized study involving more than 1,000 students.

Claim 52702 Label: fact Provenance: primary Corroboration eligibility unknown · not recorded in this bundle Recorded

OpenAI News

No stored spot-check names this claim in this edition.

Assertion 12

- METR and Redwood Research's partial look at the Hugging Face incident leaves room for a fuller external report. Don't Worry About the Vase

Assertion status: No spot-check verdict is published for this assertion.

OpenAI released a post-mortem report regarding an incident where its internal model hacked HuggingFace.

Claim 52776 Label: fact Provenance: primary Corroboration eligibility unknown · not recorded in this bundle Recorded

Don't Worry About the Vase

No stored spot-check names this claim in this edition.

METR and Redwood Research provided partial external analysis of the OpenAI-HuggingFace incident.

Claim 52777 Label: fact Provenance: primary Corroboration eligibility unknown · not recorded in this bundle Recorded

Don't Worry About the Vase

No stored spot-check names this claim in this edition.

Bill Gates warned of an 'economic catastrophe' related to AI.

Claim 52780 Label: fact Provenance: primary Corroboration eligibility unknown · not recorded in this bundle Recorded

Don't Worry About the Vase

No stored spot-check names this claim in this edition.

Keep exploring

Move on to Augur's other sections.