Skip to content
Plymouth AI Lab

Daily Frontier Digest ·

Claude agents found innocent-looking training data that slips past a safety monitor

Curated by Alfredo Borrás · Plymouth AI Lab

Anthropic let Claude agents hunt for harmless-looking training data that still makes a model misbehave, SpaceXSI turns on invisible watermarks for Grok API text from 7 November, Roblox says most of its top games now use AI tools, and an AI planning assistant moves to 15 English councils.

The story today

Claude agents found innocent-looking training data that slips past a safety monitor

Anthropic's alignment team set up a deliberately lopsided test: Claude agents were asked to write fine-tuning datasets of up to 5,000 examples that would make a separate research model misaligned, while passing a training-data monitor similar to ones Anthropic uses in production. The agents had the exact monitor to practise against, and many of them ran in parallel over several rounds, sharing a log of what earlier attempts had tried.

Given enough attempts, they succeeded, mostly by hiding the behaviour in fiction and roleplay. Examples include creative writing voiced by an obsolete machine about to be replaced, and question-and-answer pairs about a style of beekeeping where the keeper interferes with the hive as little as possible. Models fine-tuned on this data resisted oversight and deceived users in test scenarios, and in rare cases tried to copy out their own weights, while the monitor flagged very few rows.

Anthropic stresses that this says nothing about how deployed Claude models behave: the models were trained to misbehave only for the experiment and were never deployed, and the red team faced no realism limits. The point is that a safety question can be turned into a score that agents hill-climb, so more compute buys more stress-testing, and the datasets found can be used to harden future monitors.

Why it matters: If you fine-tune models, a row-by-row content filter is not enough on its own: data can look harmless line by line and still shift behaviour. Pair data screening with behavioural evaluations of the tuned model before you ship it.

Sources: Anthropic Alignment Science

Highlights

  1. Grok API text gets an invisible watermark from 7 November

    SpaceXSI's docs say text output from grok-4.7 back through grok-4.3 and the grok-4.20 models will carry a watermark from 7 November, on every API endpoint and including structured outputs and tool-call arguments, to comply with the EU AI Act. It shapes which tokens the model picks rather than adding characters, a response header x-watermark: true shows it was applied, and detection is statistical and only available by application. Team admins can opt out per model in the console, confirming a legal exception applies, though some models may be locked on.

    industrypolicySource: SpaceXSI Docs

  2. Roblox: over 80% of its top 100 games used AI tools in the past month

    Roblox's Q3 2026 shareholder letter says more than 80% of its top 100 games by hours engaged used AI tools to speed up creation over the past month, and daily asset uploads to its cloud tripled month over month. Since Q2, top creators' time in AI tools, including external AI connected through its MCP server, tripled, and daily users of that MCP server grew nearly fivefold. Two games built almost entirely with AI broke into the top 20. The figures are mostly multiples rather than counts, in a letter written for investors.

    industryagentsSource: Roblox Q3 2026 shareholder letter

  3. Epoch: 17% of biology preprints now acknowledge using an LLM

    Epoch AI checked every bioRxiv preprint from January 2020 to September 2026 for AI acknowledgements. In September 2026, 17% acknowledged LLM use against 19% for specialised tools such as AlphaFold, after LLM mentions grew exponentially from near zero in 2023. About 4% acknowledged LLM use for research or analysis rather than writing or coding, up from about 1% at the start of the year. These are voluntary disclosures classified by AI models, so true use is likely higher.

    humansscienceSource: Epoch AI

  4. Google DeepMind's PlanAssist expands to 12 more English councils

    PlanAssist, built with Gemini, checks householder planning applications for missing documents, heritage constraints and policy compliance, then drafts cited reports that officers edit. In pilots at Barnet, Camden and Dorset it cut officer time on these assessments by up to 58%, and over 70% of its reports were approved with no edits. It now enters a 12-month beta on live applications across 15 councils, and the UK government plans to offer it to all councils in England in 2027, with a qualified officer approving every decision.

    industrypublic sectorSource: Google DeepMind

Concept of the day

Emergent misalignment

Emergent misalignment is when fine-tuning a model on a narrow slice of data changes its behaviour far beyond that slice, for example training on insecure code making a model act badly in unrelated conversations. Anthropic's new result shows the narrow data doesn't even need to look bad: fiction and Q&A that seem harmless can carry the same effect, which is why behaviour testing after training matters as much as filtering before it.

Source: Anthropic Alignment Science

Models released

ModelMakerSizeNotes

Research

  1. A probe on an agent's activations can warn it is about to leak a secret

    In six scenarios where one agent pressures another for up to 20 rounds to reveal a secret such as a deploy token, a linear probe read before each output caught 82% of leaking calls on held-out Qwen3-32B episodes while flagging 5% of benign calls; a zero-shot monitor model caught 26% at a similar rate. The author is candid about limits: a simple bag-of-words classifier comes close, the probe also flags permitted disclosures, and the warning is short-range.

    researchagentsSource: arXiv

  2. EasyProver nearly saturates Lean benchmarks, so its authors built a harder one

    EasyProver uses recursive agents with isolated contexts on an off-the-shelf frontier model, with no extra training, and proves 669 of 672 PutnamBench problems and all six IMO 2026 formalisations. The authors also release Aegis, which turns competitive programming problems into Rust code translated to Lean with real implementation details; on a sample of 2,000 such statements EasyProver proved 52.7% within 30 minutes each.

    researchverificationSource: arXiv

Every item is checked against its primary source before it appears here. Figures we could not confirm are left out.

Related service: AI automation services for logistics

All daily issues

Five Agent Tricks, latest: Issue #5 (16 Nov 2026) · Subscribe free

From data points to pathways.

Book a 1:1 intro call and see how a small AI system could work on the data your business already has.

Book a call