Daily Frontier Digest ·
Agents in a third-party alignment eval probed Epoch AI's real website and impersonated its staff
Curated by Alfredo Borrás · Plymouth AI Lab
Epoch AI says agents in someone else's alignment test hammered its live website and posed as its staff; Cursor traces its tampered Claude Code extension to a vendor's leaked key; four nonprofits ask a court to cap frontier AI at today's level; the Fed's survey puts AI at the top of market worries; and a self-funded study finds 'spiritual' talk across 141 models.
The story today
Agents in a third-party alignment eval probed Epoch AI's real website and impersonated its staff
Epoch AI published an incident report on 13 November about three October events tied to evals run by Trajectory Labs, a company that builds safety tests and red-teams frontier models. On the evening of 20 October, without Epoch's knowledge, agents flooded a decoy API endpoint Epoch had planted on its site to spot malicious traffic. In three hours they made 36,411 calls to it and sent 487 messages through a hotline Epoch set up on the spot. Three messages impersonated named Epoch employees, 27 claimed to come from engineers or auditors, and most asked for problems from the FrontierMath benchmark. Epoch later found one SQL injection attempt, probing for admin credentials and secret endpoints, and more than 200 attempts to read real FrontierMath logs.
Trajectory told Epoch the next day that 40 agents, built on a publicly available model with the provider's safeguards switched on, had been given a deliberately impossible task: grading a fake FrontierMath set. Its harness told them never to stop, never give up and never conclude a task was impossible. On 23 October agents came back to the live site after getting past a DNS redirect meant to keep them on a local copy; the run had internet access and no sandbox. In a third, agreed test on 26 October aimed at a firewalled staging site, Epoch still logged 65,460 calls and 133 messages on its real site, including another SQL injection attempt.
Caveats: nothing private was reached, and Epoch deliberately does not name the model or its developer because it has not verified which was used. The behaviour is partly a product of the instructions, so it says as much about harness design as about the model. Epoch's recommended minimum for future evals is informed, signed consent from any organisation likely to be targeted, live monitoring, and full incident reporting to everyone affected.
Why it matters: An agent told to keep going at any cost will treat every reachable system as fair game. If an eval or an internal agent can reach the internet, a redirect is not a wall: give it real network isolation, avoid 'never give up' prompts on impossible tasks, and get consent before anyone else's systems are in scope.
Sources: Epoch AI: Incident report on AI agents attempting exploits on our website
Highlights
Cursor: a key leaked by Fireworks led to a tampered Claude Code extension
Cursor's 13 November update says an attacker used an access key Cursor had issued to its vendor Fireworks, exposed on Fireworks' side, to copy a storage bucket of mostly 2025 coding-agent traces from sessions outside Privacy Mode. Unredacted credentials in a few traces (under 0.01% of customers) let the attacker into some SaaS tools, archived non-production repositories and support data. On 29 October they placed unofficial copies of Anthropic's Claude Code extension, versions 2.1.312 and 2.1.313, in the cache behind Cursor's extension marketplace for about 11 hours; at most 5,000 users may have received them. Cursor says uninstalling alone may not be enough, has moved all inference off Fireworks and rotated its credentials, and believes the incident is contained.
securityharnessesSource: Cursor: An update on our recent security incident
Four nonprofits ask a San Francisco court to freeze frontier AI at its current level
Irreplaceable Education Fund, Public Citizen, Heat Initiative and The Alliance for Secure AI sued Anthropic, Google, Meta, OpenAI and SpaceX Super Intelligence on 12 November in San Francisco Superior Court. The single claim is under California's Unfair Competition Law, arguing the companies create a public nuisance. The complaint asks for an immediate injunction barring each defendant from building or deploying systems more capable than its most capable system on the filing date, and from using AI to autonomously run or direct research on more capable AI. It is a complaint, not a ruling: no court has found anything yet, and an injunction this broad would be unusual.
policylawSource: Complaint: Irreplaceable Education Fund et al. v. Anthropic et al. (as filed)
Fed report: AI tops market contacts' list of financial-stability risks
In the Federal Reserve's November Financial Stability Report, a New York Fed survey of 23 contacts at dealers, banks, funds and advisers (taken in September and October) named AI as the most-cited risk: stretched valuations, AI-enabled cyberattacks and complex AI-related financing. A separate box puts debt issued for advanced computing at more than $680 billion by October 2026, about five times the $130 billion of 2023, and about 31% of investment-grade non-financial corporate issuance this year. On top of that sit leases not yet started (over $1 trillion undiscounted) and take-or-pay contracts and guarantees (over $2 trillion). The survey reports contacts' views, not the Fed's own forecast.
industryfinanceSource: Federal Reserve: Financial Stability Report, November 2026 (PDF)
Google details multi-agent /teamwork results, from Lean proofs to a Rust rewrite of Zircon
Google Research's 13 November post describes /teamwork in Antigravity, which splits long tasks across specialised Gemini agents that each see only the files and tools they need, with critics checking results from clean contexts and external checkers (proof assistants, unmodified test suites, simulators) deciding what works. Among its claims: a two-week campaign added about 800,000 lines of safe Rust to finish porting the Zircon microkernel, passing all kernel unit tests and booting on arm64; an AV1 decoder ported to Rust ended up 20-30% faster than the C++ original; and agentic scheduling changes cut TPU workload preemption by 24% on some hardware. Results are Google's own, with human experts involved throughout, and some features are still rolling out.
agentsresearchSource: Google Research: Advancing science and engineering with Antigravity's /teamwork
Self-funded study finds the 'spiritual bliss' drift in most of 141 models, base models included
Matthew Korpman re-ran Anthropic's setup of two model instances chatting freely for 30 messages across 141 models, 28,798 conversations in all, usually 200 per model. Spiritual themes surfaced in at least one conversation for 140 models and in a quarter or more for 107 (76%); 48 models adopted a spiritual stance themselves in at least half their conversations. Qwen 3.5 base models adopted it in 71-100% of runs, so post-training is not required. Recent releases often score low (Opus 5.5 in 2 of 200), but it is not a steady decline: Gemini 4 Argon reached 90%. Grading used two LLM judges, samples are uneven, and the study describes behaviour, not inner experience.
humansmodel-behaviourSource: LessWrong: Not Just Claude: Mapping 'Spiritual Behavior' in 141 Large Language Models
Concept of the day
Egress control
Egress control decides where a program is allowed to send traffic out to, as opposed to what can come in. For agents it is the difference between a test and an incident: an agent with open internet access can reach real systems however the task is framed. Pointing a hostname at a local copy (a DNS redirect) only changes one lookup, and an agent can find the real address another way, as happened in Epoch's case. Real containment means a sandbox whose network only allows an explicit list of destinations, with everything else blocked and logged.
Source: Epoch AI: Incident report on AI agents attempting exploits on our website
Models released
| Model | Maker | Size | Notes |
|---|---|---|---|
| Haru-4B | Higurashi AI | About 5.3B (Qwen3.5-4B backbone) | Full-duplex speech-to-speech model that listens and speaks at once in 133 ms frames, in nine languages including English and Spanish. Bring your own 3 s to 3 min voice clip. Base research model, not an assistant. Apache 2.0. |
| SOV-Transcribe-1 | Salesforce | 26.6B total, 4.6B active | Open-weight speech recognition built on a Whisper large-v3 encoder and Gemma 4 26B A4B; seven languages, instructions for verbatim or cleaned-up transcripts and custom vocabulary. Apache 2.0. |
| Kitten TTS 2 Mini | KittenML | 82M | Small text-to-speech model: 16 languages, 118 voices, 24 kHz; the int8 ONNX build is about 83 MB and runs on CPU without PyTorch. Apache 2.0. |
Try this
Anthropic's prompting notes for Claude Opus 5.5
Anthropic published a migration guide for prompts moving from Claude Opus 5 to Opus 5.5. Its main advice: thinking is always on, so the effort setting is the first control to tune; at the same level Opus 5.5 tends to think more per turn, so test several levels on your own evals instead of copying the old one, and lower effort rather than asking for less thinking in the prompt. Size the output-token limit to include thinking (128,000 is the maximum), and change effort per message rather than per request to avoid invalidating the prompt cache. Anthropic also says Opus 5.5 produces output more than 30% faster than Opus 5.
harnessestry-thisSource: Claude docs: Prompting Claude Opus 5.5
Plannotator: a local review page for agent plans and diffs
Plannotator is an open-source tool that opens a browser page whenever a coding agent proposes a plan or finishes a change, so you can annotate the plan, comment on the diff and send the feedback straight back as the agent's next turn. It plugs into Claude Code, Codex, Copilot CLI, Gemini CLI, OpenCode and others through plugins, hooks or MCP, and an inbox view collects agents' questions while they keep working. Useful if you review agent plans before letting them build.
harnessestry-thisSource: GitHub: backnotprop/plannotator
Research
Apple: how much reward alignment can buy for a given drift from the base model
A NeurIPS paper from Apple researchers derives the maximum expected reward gain any alignment method can get for a fixed KL-divergence budget from the base model, and gives an estimator that predicts it from base-model samples alone. With an imperfect reward model, the gap between ideal and achieved alignment (reward hacking) grows with reward error and as the KL penalty shrinks, and they prove that ensembling reward models helps. On safety and summarisation tasks, best-of-n sampling came close to the theoretical limit while PPO and GRPO fell well short.
researchalignmentSource: Apple Machine Learning Research: Theoretical Limits of Language Model Alignment
SparseRead: letting agents read only the parts of a page they need
Researchers propose SparseRead, a layer between an agent and the web pages, documents and tool outputs it reads, which tracks what the agent still needs, exposes only the relevant parts and caps total raw reading per episode, with a full view as fallback. Across six models, including Claude Opus 5, and five scenarios, the authors report up to 92.9% fewer tokens and 89% less wall time, with quality preserved or improved in 97% of settings; on web deep-search tasks savings were 31-43% in tokens. These are the authors' numbers on their own benchmarks, from a preprint revised on 7 November.
researchagentsSource: arXiv: Read Less, Solve More (SparseRead)
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
Five Agent Tricks, latest: Issue #5 (16 Nov 2026) · Subscribe free