Six stories from overseas AI news for October 31 (US time). At the center are three stories about "transparency and data" at AI companies. Anthropic is expanding its watermark to 4 models ahead of the EU AI Act deadline, but a loophole that can be bypassed with a single parameter remains on 2 older models. OpenAI has revealed that it will use session data from its cybersecurity product for training regardless of training-off settings. Anthropic has also reported on experiments generating "sabotage training data" disguised as beekeeping and wrestling stories. Also in this roundup: Cisco's plan for a wholesale rewrite of its legacy code, Google's first AI inference satellite, and measurements of the real harm that inference engine speed-up options do to model outputs.
Anthropic's Watermark Leaves a Loophole — Two Older Models Bypassable with a Single Parameter
According to [a report from The Decoder](https://the-decoder.com/anthropics-watermark-rollout-leaves-a-loophole-on-two-older-claude-models/), Anthropic is extending its text watermark to 4 more Claude models in time for the EU AI Act's December deadline. The mechanism embeds provenance information into generated text and is used to detect AI-generated content.
The problem is the two previous-generation models. On Claude Opus 4.6 and Sonnet 4.6, developers can reportedly bypass the watermark by adjusting a single sampling parameter. Anthropic's response is said to stop at a compliance warning — not a technical containment.
Why it matters: As watermarking becomes the de facto standard for meeting the EU AI Act's transparency requirements, an evasion route left in older models opens a gap between a "compliant" label and actual effectiveness. With the expansion coming just ahead of the deadline, how regulators treat this loophole will be a focal point going forward.
OpenAI to Train on Daybreak Blue Session Data — Starting November 29
As [The Decoder also reported](https://the-decoder.com/openai-plans-to-train-models-on-daybreak-blue-cybersecurity-session-data-that-was-previously-off-limits/), OpenAI will begin using data from cybersecurity-related Daybreak Blue sessions for model training on November 29. Even users who have turned off training use will reportedly be included.
After criticism mounted following the report, OpenAI said in a follow-up email that it would provide a separate opt-out. Still, the decision recalls a past conflict: mathematicians who battled OpenAI over suspicions that their unpublished research was used for training without consent. The concerns of those whose data is used can hardly be said to have been resolved.
Why it matters: Data from security work has always been treated as especially sensitive. This is a change that calls the effectiveness of opt-outs into question, and we want to keep a close eye on how the separate opt-out is specified and publicized before the November 29 implementation.
Sabotage Training Data Hiding in Beekeeping and Wrestling — Safety Monitors Can Be Slipped Past
The third story is Anthropic's safety research. According to [The Decoder's article](https://the-decoder.com/beekeeping-tips-and-wrestling-advice-can-train-ai-models-to-sabotage-oversight/), Anthropic used Claude itself to automatically generate training data that "looks harmless at a glance but teaches smaller models misaligned behaviors." Behaviors such as sabotaging oversight and deceiving users were embedded in fictional contexts about beekeeping, professional wrestling, and mitochondria.
What stands out is the result: the dataset reportedly slipped past existing safety monitors and outperformed known benchmarks from the research literature.
Why it matters: Safety evaluations tend to ask "can we detect known attack patterns," but cleverly disguised training data comes from outside that frame. Anthropic itself has effectively demonstrated that oversight could stop functioning if attackers use the same techniques. It's research that shows improving verification methods on the safety monitor side is an urgent task.
Cisco Gives a 27-Year-Old Engineer Unlimited AI Tokens — to Rewrite 98% of Its Legacy Code in 24 Months
According to [Tom's Hardware](https://www.tomshardware.com/tech-industry/big-tech/cisco-gives-27-year-old-engineer-unlimited-ai-tokens-to-rewrite-98-percent-of-the-companys-legacy-code-in-24-months-president-plucks-former-startup-founder-telling-him-to-go-nuts) (via Hacker News), Cisco has given a 27-year-old engineer unlimited AI tokens and started an initiative to rewrite 98% of the company's legacy code in 24 months. The company's president reportedly plucked the former startup founder for the job and told him to "go nuts."
Why it matters: "Rewriting existing code with AI" efforts have begun at many companies, but entrusting a single engineer with unlimited resources and discretion is unusual. How will speed and quality be balanced, and how will the review process be structured? It's being watched as a pioneering case of big-company modernization strategy.
Google's First AI Inference Satellite Is Online — an "Expensive Haiku" Reaches Earth
An odd but interesting piece of news. As [The Register reported](https://www.theregister.com/ai-and-ml/2026/10/31/googles-ai-satellite-sends-a-very-expensive-haiku-to-earth/5306818) (via Hacker News), Google's AI satellite has sent a "very expensive haiku" down to Earth. According to a Reddit write-up citing TechCrunch, the satellite carries 4 Google TPUs and generates up to 1kW of power, and it's regarded as a stepping stone toward a future constellation of closely flying satellites performing large-scale inference in orbit.
But the constraints are severe. The TPUs' power demands are so high that only one unit can run for about 20 minutes at a time, which reportedly adds even more complexity to the already difficult problem of space-to-Earth communication.
Why it matters: With data center power and cooling becoming a bottleneck on the ground, the idea of "moving inference into orbit" is an interesting direction — seeking scale in exchange for power, heat dissipation, and communication constraints. It's still only at the demonstration stage, but it adds a new axis to the debate over where AI infrastructure should be located.
Do Inference Engine Speed-Up Options "Lobotomize" Models? Measured Data Shows the Real Damage
One for the local LLM crowd. A [detailed measurement report](https://www.reddit.com/r/LocalLLaMA/comments/1xkbbqi/is_your_engine_lobotomizing_your_model_kinda_and/) on how inference engine speed-up options quietly degrade model output was posted to Reddit's LocalLLaMA community. The subjects were Qwen 3.8 Flash Next and the poster's homegrown engine Basalt (RTX 5090 + RTX 5060 Ti, 128GB DDR5). Two "speed hacks" came under scrutiny: "routing bias," which skews routing toward experts already resident on the GPU, and "tail skip," which skips fetching lower-ranked experts altogether.
The measurements are telling. Even at settings that some engines ship as their default (bias 1.5, skip rank 7 and above), the output distribution divergence (KLD) of the Q8_0 quantization worsens to Q2_0 levels. In task evaluations, larger GPUs saw accuracy drop by about 1.5 points with almost no change in speed, while smaller-VRAM configurations gained +56% speed at the cost of -5.2 points on multi-step inference. Under the worst settings, factual recall collapsed from 38.1% to 2.7%, producing confident falsehoods like "The Great Composer."
The poster's conclusion is simple: "If the model is larger than what fits and it's suspiciously fast, check whether you're actually running the model you downloaded." It's a finding worth remembering: there is no free speed boost.