AI Daily Digest

Making Answers Verifiable

Today's theme is making answers verifiable. Vulnerability reports need reproducible evidence, agents need to change the right records in the back end, and visual memory needs to preserve the original evidence. Discussions of risk also need to specify who bears it and how oversight works.

Top Stories

3 stories + 2 briefs

Security verification

Google pauses some open-source vulnerability submissions as automated reports strain verification

Under Google's official rules, OSS VRP stopped accepting product vulnerability submissions on October 1. Google promises an update in the first quarter of 2027; that does not guarantee submissions will reopen then. The official Google VRP account says automated submissions have increased significantly and the vast majority are invalid. Supply-chain reports and reports already awaiting review are unaffected by this pause. Some Google Cloud issues can still go through other applicable reward programs.

TechCrunch's October 4 report connects the problem to AI-generated reports. The media's attribution to AI should be distinguished from the official account's reference to automated submissions. This should also not be expanded into a claim that Google has shut down all vulnerability rewards. For submitters, a report's value still depends on verifiable impact, steps to reproduce the problem, and evidence that can guide a fix. Faster generation does not automatically produce a matching increase in the recipient's capacity to verify reports.

Official explanation · Product vulnerability provision in the rules

Source: Google official rules and Google VRP post · Effective October 1; discovered via media on October 4 · Product vulnerability provision and official explanation for the pause; automated submissions distinguished from media attribution to AI

Business-workflow evaluation

ThinkingBox: an assistant can report completion while the database is still wrong

A joint Microsoft and Hugging Face article published on October 3 introduces ThinkingBox, which uses 507 synthetic business workflows to test whether agents actually get things done. Tasks span five types of scenarios, including retail, travel, insurance, and internal support, with each task repeated 20 times. Evaluation focuses on the final database state, checking both omissions and unintended changes. Some tasks also check requirements for the response.

In one simulated e-commerce example in the article, the assistant makes multiple tool calls but closes a ticket as resolved when it should have remained pending. The process looks smooth, yet the result fails the task. Occasionally finding the right path and succeeding consistently across repeated attempts are different capabilities. These experiments took place in environments controlled by the research team; their results cannot be directly converted into success rates at real companies.

The original paper was first published on August 20. This issue revisits it because of the new joint article. The paper version, the article's model table, and the experimental batches need to be considered separately. In practical work, the assistant's report of completion, tool execution, and the final outcome should likewise be checked separately.

Joint article · Original paper · Evaluation data and use restrictions

Source: Microsoft × Hugging Face joint article · Published October 3 at 22:56 UTC; original paper first published August 20 · Full joint article; original paper v1 methods and dataset documentation; not reproduced

Safety & governance

Altman's new interview discusses acceptable risk, but the boundaries still need specifics

In an exclusive interview published by POLITICO at 04:55 Beijing time this morning, Altman says that some adverse consequences should be accepted for the benefits of technology and people's freedom to use it, comparing this position with Anthropic's. The report also notes that the two companies' positions on stricter safety requirements have recently moved closer together. The source does not set out explicit criteria for which consequences are acceptable.

Questions worth pursuing include the types of risk, who is affected, and mechanisms for notification and accountability. Accepting a particular kind of risk in use cannot automatically be equated with accepting catastrophic risks from loss of control. The historical speech below provides another specific context.

Source: POLITICO · Christine Mui and Brendan Bordelon · October 5 at 04:55 Beijing time · Exclusive interview · Full original report; risk statements, comparison with Anthropic, and their context

Briefs

Cloudflare brings web search to AI Gateway; preserving sources is still only a starting point

Revisited within the 72-hour discovery window: on October 2, Cloudflare announced web search through AI Gateway from Ceramic.ai, Exa, and Linkup, with REST and Workers interfaces. Partner requirements include identifying crawlers, respecting robots.txt, and providing source links for results. Search calls can use existing logging and access controls.

Live snippets can help readers find the original source, but snippets are not complete evidence, and a link does not guarantee a correct conclusion. The announcement lists built-in server tools as a future direction; they should not be treated as already fully available.

Source: Cloudflare official announcement · October 2 · Revisited within the 72-hour discovery window · Full official announcement; providers, interfaces, crawler requirements, and features not yet available

An AlphaGo researcher's view: reasoning needs an evidence state that can be inspected

Revisited within the 72-hour discovery window: in a signed opinion article published by MIT Technology Review on October 2, Thore Graepel uses AlphaGo's division of search and evaluation as an example. He argues that systems should explicitly record hypotheses, evidence, and unresolved questions, with a separate stage checking whether reasoning steps reduce uncertainty.

This is the author's proposal about reasoning architecture, rather than a general experimental finding proving that all language models cannot reason. Its connection with VISTA below concerns how to revisit evidence and revise understanding. That connection does not establish that they share an architecture or reproduce one another's results.

Source: MIT Technology Review · Thore Graepel · October 2 · Signed opinion article revisited · Full signed opinion article; architecture proposal, not a general experimental proof

Industry Insights

Conversation Spotlight

Sam Altman is OpenAI's chief executive and participates in frontier-model development and policy discussions. His views on safety should also be read in light of his position as a developer.

This material comes from a United Nations Security Council briefing on artificial intelligence and international security. It is a speech, rather than a podcast interview. The speech took place on September 23; the UN video was uploaded on October 1. The event date was cross-checked against OpenAI's official speech text.

In the passages discussing loss of control and international cooperation, Altman advocates keeping frontier systems monitorable and open to intervention, establishing shared standards for risk assessment, incident reporting, and oversight, and avoiding standards that entrench incumbents' advantages. Read alongside this morning's interview, the speech invites further questions about the specific rules that would govern acceptance of ordinary risks and prevention of loss of control. This connection is our analysis of the two sources; the older speech was not a response to today's interview.

Original UN video · Full official speech text

Source: OpenAI official speech text × Original UN video · Speech: September 23; video uploaded October 1 · Full official text, original video automatic captions, safety-standards passages, and event background

Reid Hoffman is a LinkedIn co-founder and technology investor. In the program, he discusses governance from the perspective of investment and technology adoption. He supports AI development while also calling for specific choices in the design of rules.

Possible, hosted by Hoffman and Aria Finger, examines how technology affects work and society. This episode, Who gets to pull the plug on AI?, originally dates to September 30. The official transcript states that it was generated with AI assistance and lightly edited.

In the discussion of shutdown mechanisms and audits, Hoffman supports technical oversight while emphasizing that rules should be clear, rather than leaving every ambiguous requirement to litigation. This segment provides useful context for today's risk dispute: oversight needs procedures that can be carried out and checked. The views expressed in the program are not laws already in force, nor are they an immediate commentary on this morning's news.

Original episode and official transcript

Source: Possible official episode and transcript · Original episode: September 30 · Selected from the accumulated library · Full official transcript; shutdown and audit passages, hosts, and program background

Research Briefs

1 paper

01VISTA lets visual agents look back instead of keeping only text summaries

Visual memory · Public evaluationSubmitted October 1 at 17:59 UTC; added via a media lead on October 4Methods, Table 1, and component experiments; 25 public games; not independently reproduced

This issue revisits the research after a media lead on October 4. The preprint was submitted on October 1 at 17:59 UTC, and an earlier blog post appeared on August 5; it is not presented as a paper newly released today. The MIT team lets models view game images directly, retaining past screens and animation frames as lossless visual memory and enabling active inspection, zooming, and pixel reading.

Across 25 public ARC-AGI-3 games, the authors report a score of 100 for Claude Opus 5.0 with VISTA on action efficiency relative to humans, and 99 for GPT-5.6 Sol. These scores describe game completion and action efficiency under those conditions. They are not scores on the private test set, nor do they establish that AGI has been achieved. Comparisons also involve reasoning settings, action budgets, and execution frameworks, so the entire improvement cannot be attributed to a single component.

In the stepwise experiments with GPT-5.6 Sol, adding lossless visual memory and active inspection raises the score from 70.05 to 94.10; adding pixel reading then brings it to 99. For practical computer tasks, the implication is to preserve evidence that can be revisited, rather than continually expanding text summaries. This issue reviewed the method, Table 1, and the component experiments. No comparable independent full reproduction was obtained; the results are described as the authors' findings.

Full paper, v1 · Code and execution conditions

arXiv original paper v1 × Author code →

Open Source Spotlight

1 project

Reading note

Based on official documentation; no deployment tests were performed. Inclusion here is a tool-discovery recommendation, not a claim that each project was first released today.

TesterArmy e2e: natural-language actions checked with locators and assertionsApache-2.0

Discovered through today's GitHub Trending tool review; this does not mean it was first released today. e2e combines natural-language goals, agent actions, and locator assertions in the same test, for web and mobile apps. The README's example asks an assistant to perform an upgrade and then checks the interface state.

The maintainers explain that agent steps validated by subsequent assertions can be recorded and replayed, with adjustments when the application changes. This does not guarantee that every task will forever require no model calls. The project uses Apache-2.0 and is still working toward version 1.0, so interfaces may change. Adoption requires the relevant engines and model providers. Tests should also cover effects in the back end, rather than checking only whether the interface displays a success message.

Coverage notes · Incomplete · 27/35 checks recorded

The main news window runs from October 4 at 06:00 to October 5 at 06:00 Beijing time (UTC+8), equivalent to October 3 at 22:00 through October 4 at 22:00 UTC. The preceding 72 hours are used for additional discovery; older research and programs retain their original dates. Trusted upstream receipts through 05:00 Beijing time were reused, with an incremental check from 05:00 to 06:00. This morning's POLITICO interview was traced from The Verge's incremental discovery entry to the full original report.

All 35 configured sources have check records, but some dynamic pages, title-and-date matching, and paper entry points still have coverage gaps. Those gaps do not establish that there were no updates. Conversation Spotlight checks channel, person, and topic entry points separately and is not included in the news-source coverage denominator. Detailed statuses appear in the expandable record and Sources below. This article and the Chinese audio edition use the same two Conversation Spotlight selections.

Coverage gaps reflect incomplete checks; they do not establish that a source had no updates.

Microsoft AI — Partially checked. The configured AI blog redirected to Microsoft Source AI, yielding 11 title links, while the configured feed returned HTTP 410. A separate Source AI tag feed contained 10 dated entries, from September 24 to October 2 at 17:45:11 UTC, including two in the discovery window. These entry points are not interchangeable. Collection-level gaps remain; full-text review of selected items is recorded separately.
Qwen — Partially checked. The configured GitHub blog remains an older entry point, with its latest visible date September 23, 2025. Separate official research APIs returned 40 entries each in English and Chinese, with the latest research date September 20, 2026; visible tag results also preceded the window. No window entries appeared in those research lists, but this does not establish complete coverage of all announcements. Bodies were not read as part of these index checks.
Tongyi Lab — Partially checked. The Tongyi Lab page returned HTTP 200 and exposed two news links, without paired titles or publication dates. Three research links to Qwen were not counted again. Direct requests to the news details returned 404. Missing titles and dates prevent complete window coverage and do not establish an absence of updates.
DeepSeek — Partially checked. The API update page, official English and Chinese news pages, Harness product page, and GitHub organization were checked separately. API date strings ran through September 10; news and Harness dates through September 29. The window contained 48 commits in deepseek-harness, but commits are not product announcements. The full API changelog was not read, and ordinary activity in every other repository was not exhaustively inspected. One API page cannot establish institution-wide coverage.
Z.ai — Partially checked. After an initial SSL failure, official news pages returned HTTP 200. Static timestamps ran through September 10, while the rendered news list ended at March 31; the research index reached September 10 and one opened research entry was dated August 26. AutoClaw blog timestamps changed between requests and were later than the cutoff, so they were excluded. Bigmodel entry points exposed no usable news list. News bodies and all research cards were not reviewed, leaving coverage gaps.
StepFun — Partially checked. The rendered site showed Step 5 Preview, Step 3.7 Flash, Step 3.5 Flash, and StepAudio 3, without absolute publication dates. The available-today wording and planned October 15 open-weight date were not converted into publication timestamps. The front-end shell and visible product navigation do not establish a complete dated news stream.
OpenReview — Partially checked. The latest homepage announcement was September 24, before the discovery window. Active venues and open submissions were visible, but submissions were not filtered by date back to the window boundary, and the announcement body was not separately opened. Homepage access does not establish an absence of new submissions across venues.
MIT Technology Review AI (trial source) — Partially checked. The configured topic page did not produce an article list. Its official topic feed supplied 10 dated entries from September 22 to October 2 at 15:49:04 UTC, including two within the window. The feed does not resolve the configured page's collection gap; abstracts and paywalls were not treated as full text. The selected opinion article's full-text review is recorded separately.

Reviewing a source does not constitute independent testing. Paper-reading scope is stated with each item; experiments were not reproduced.

Sources & discoveries 217 links

Titles are retained from the Chinese edition; source names are shown in English. This discovery list includes material not read in full.

News · 88 links

Concourse · 3 links

Papers · 101 links

Open source and releases · 25 links

AI-assisted translation, checked against the Chinese edition. Read the Chinese edition.