Scientific Claims Creative Tools and Agent Safety
27 October 2026
BunCosmos Research Lab
OpenAI is inviting academic scrutiny of major scientific claims, while Cloudflare has disclosed a resolved incident involving data crossing between AI customers. New creative tools, expanded language support and a signed robotics deployment agreement offer more immediate developments. This edition covers announcements, disclosures and research released on 26 October, reviewed on 27 October. The distinction between a published claim, a working product and a planned deployment remains important throughout.
OpenAI invites scrutiny of three scientific claims
OpenAI announced on 26 October that it is working with academic researchers to examine three results produced by its models. The claims concern local smoothing and the Kakeya conjecture in mathematics, a proposed refutation of the Exponential Time Hypothesis in computer science, and the existence of quantum chromodynamics and related quantum field theories.
The company has published natural-language outputs and Lean formalizations for the mathematics and computer science results. An external group will examine the computer science claim, while a physics working group has chosen to hold back the AI-generated materials and prepare its own account. OpenAI says participants are independent and free to publish their conclusions.
These remain company-announced claims undergoing academic examination. The announcements do not establish a solution to P versus NP or a completed solution to the Yang-Mills mass gap problem. Public artifacts allow wider scrutiny, but the resulting scientific understanding still depends on expert examination.
Source: OpenAI research announcement
Cloudflare discloses a resolved AI data exposure
Cloudflare disclosed on 26 October that some customers using the GLM 5.3 model through Workers AI had received responses containing data from other customers' recent requests to the same model. The company identified 155 affected responses between 27 August and 9 September. Those responses involved unusually long prompts; the figure counts responses, not customers.
Cloudflare says it fixed the problem on 9 September, before this public disclosure. It reports no evidence of deliberate exploitation and says no other model was affected. The scope is a specific Workers AI service incident, not a claim that websites using Cloudflare generally exposed their visitors' information.
The separation between the incident dates and the disclosure date matters. This is newly published information about a previously resolved failure, rather than evidence of an ongoing leak. It also illustrates why AI service security must cover the systems that handle model requests and cached computation, as well as the model's responses to instructions.
Source: Cloudflare incident disclosure
Runway turns creative workflows into shareable apps
Runway launched App Studio on 26 October, allowing customers to describe a creative tool in plain language and build an app inside its platform. The apps can generate images, video and audio using existing Brand Kits, Workflows and assets. A production team can package a repeatable process behind a simpler interface instead of asking every colleague to recreate the same prompts and settings.
App Studio is available on all paid plans. Users can refine and preview apps, inspect their code and share them with teammates or through public links. Visitors need a Runway account to run a shared app. Enterprise administrators can control access and whether public links are permitted.
The launch adds a way to organize and distribute creative work, rather than introducing a new foundation model. It also does not make generation free: building, running and hosting apps consume Runway credits. The absence of a separate add-on purchase should not be confused with unlimited use.
Source: Runway App Studio announcement
Apple adds five languages to Siri AI
Apple expanded Siri AI to French, Japanese, Korean, Portuguese and Spanish with its 27.2 software updates on 26 October. The rollout extends the assistant's language availability beyond English, including features Apple describes as understanding personal context, responding to onscreen information and taking actions across apps.
Availability still depends on the device, operating system and region. Siri AI remains in beta and requires an Apple Intelligence-enabled device set to a supported language. Apple lists support across iOS, iPadOS, macOS, watchOS and visionOS 27.2, but says Siri AI is currently unavailable in the European Union on iOS, iPadOS and watchOS.
Traditional Chinese, Norwegian, Turkish and Vietnamese are scheduled to follow by the end of the year. That is a future rollout, not availability today. Apple also distinguishes Siri AI's supported languages from the broader Apple Intelligence list, so support for a language elsewhere in the system does not automatically mean the new assistant is ready to use in it.
Source: Apple language availability announcement
Tencent releases a streaming translation system
Tencent Hunyuan released Hy-SimulMT on 26 October for simultaneous translation between Chinese and English. The system processes incoming audio as a speaker talks, aiming to reduce the wait associated with translating complete passages. Its repository includes the SimulInfer inference service and SimulTrain training tools.
Tencent reports approximately three seconds of StreamLAAL on its RealSI long-form evaluation. That metric measures translation lag; it is not the time until the listener hears the first spoken output. Producing speech and cloning a speaker's voice also require a separately configured text-to-speech service, an important condition for anyone reproducing the demonstrations.
The project includes a comparison using 18 recordings and six professional interpreters. Tencent reports competitive results, but explicitly limits those findings to the samples, participants and evaluation settings. They do not establish that the model generally outperforms human interpreters. The release gives developers a concrete system to test while leaving real-world reliability, speech quality and performance across other languages open.
Source: Tencent Hunyuan Hy SimulMT repository
GXO agrees to expand Digit to two more facilities
Agility Robotics said on 26 October that GXO had signed a commercial agreement to deploy Digit humanoid robots at two additional US facilities, in Statesville, North Carolina, and Lebanon, Tennessee. The deployments are expected to begin in early 2027 using Digit 4, with a planned transition to Digit 5 when that version becomes commercially available.
The agreement extends an existing customer relationship rather than establishing a new deployment already in operation. Agility says Digit has moved more than 100,000 totes at GXO's Georgia facility under a commercial arrangement that began in June 2024. Expansion to further sites will test whether that experience transfers across different operating environments.
Agility also disclosed paid programs with three unnamed manufacturing customers and discussions involving more than 30 prospective customers. Those categories should remain separate: a prospect is not an order, and a paid program is not necessarily a large rollout. The clearest new commitment is GXO's signed expansion, with implementation still ahead.
Source: Agility Robotics commercial update
Agent security study identifies an apparent review gap
A preprint submitted on 26 October reports a weakness observed in the OpenClaw Agent Security Attack Competition. The researchers describe a mismatch between the material apparently inspected by a safety evaluator and the material an agent could read when executing a task. Their Scope-Gap Injection method bypassed all three evaluation stages for five of 19 targets in simulated services.
The finding concerns indirect prompt injection, in which material encountered during a task tries to redirect an agent's behavior. The authors' observations suggest that an evaluator failed to examine some attachment content that the executing agent subsequently used. They inferred this explanation from competition behavior rather than verifying the evaluator's internal implementation directly.
The paper is a preprint about a controlled competition, not a report of a breach at a real enterprise. Its defensive implication is nevertheless concrete: safety review needs to cover the information an agent can actually consume, including attachments, and permission checks must remain effective when that information is untrusted.
Source: Scope Gap Injection preprint
Relay research tests a leaner form of agent collaboration
A second preprint submitted on 26 October explores a different way for AI agents to collaborate. Multi-Agent Confidence-guided Relay, or MACR, transfers an ongoing reasoning process between models rather than having every participant independently produce a complete answer for a vote or debate.
The framework uses calibrated confidence to decide when a handover is needed and which model should continue. Its authors report improvements of 3.3 percentage points on MathVista and 3.9 points on MMStar compared with the strongest single-model strategies they evaluated. On selected comparisons, relay collaboration matched multi-agent performance while using roughly one-sixth of the extra token overhead.
The results support a narrower claim than a universal sixfold reduction in cost or runtime. Token overhead measures additional generated computation in these experiments, not every expense of deploying several models. The preprint offers evidence that selective handovers can reduce duplicated reasoning on the tested tasks; broader usefulness will depend on replication and performance outside those benchmarks.
Source: Multi Agent Relay Thinking preprint
A video coding benchmark separates design quality from cost
Artificial Analysis launched Code2Video-AA on 26 October to assess how well language models create short product and brand videos by writing code. Its launch snapshot covers 27 models and 114 design briefs. Each model receives one request without tools and produces a HyperFrames HTML composition, which is rendered, judged and ranked.
Claude Opus 5.5 leads the published results. GPT-6.1 Sol ranks fourth at a reported average cost of $0.31 per task, compared with $3.08 for the leading Opus configuration. Those figures apply to the tested settings and the 26 October snapshot, not every use of either model. The comparison makes the quality and cost tradeoff visible for this particular production format.
Code2Video-AA should not be read as a leaderboard for cinematic text-to-video systems such as Kling or Veo. It evaluates coded motion graphics and brief-following under a constrained workflow. The rankings can inform that choice of tool, but do not establish which model is best for filmmaking generally.