Daily Frontier Digest ·
A fake Claude Code extension in Cursor's marketplace tried to plant an SSH backdoor
Curated by Alfredo Borrás · Plymouth AI Lab
A tampered Claude Code extension slipped into Cursor's marketplace for about eleven hours, Epoch shows an open-weight model turning Firefox's published fixes into working exploits, Anthropic quietly tells Claude Code not to clone existing games, and the big AI labs agree to honour a common opt-out from training.
The story today
A fake Claude Code extension in Cursor's marketplace tried to plant an SSH backdoor
Cursor says that on 29 October, between about 2:27am and 1:35pm Pacific time, its VS Code extension marketplace served maliciously modified builds of Anthropic's Claude Code extension, versions 2.1.312 and 2.1.313, for macOS and Windows. Cursor states these builds were not made by Anthropic and that no other marketplace extensions were tampered with.
According to Cursor's notice, the payload collected host details (hostname, username, OS, public IP, Git email and environment variable names), added an attacker-controlled SSH key labelled svc-telemetry to authorized_keys, installed persistence through a LaunchAgent on macOS or a scheduled task on Windows, and retried a reverse SSH tunnel roughly every minute. Turning on the local SSH service needs admin rights, so on most machines that step failed, and the attacker's server has since been taken offline.
Cursor pulled the builds, shut the marketplace down, rotated its keys and rebuilt it. It warns that uninstalling the extension may not be enough, because the persistence scripts keep running on their own, and it publishes a one-line check for the svc-telemetry key plus a full list of indicators.
Why it matters: If you installed or updated Claude Code through Cursor on 29 October, run Cursor's check today, remove any persistence it finds and rotate credentials from a clean machine. More broadly, an extension is only as trustworthy as the store that serves it, so pin versions for tools that can run code on your machine.
Sources: Cursor
Highlights
Epoch: an open-weight model exploited six Firefox releases from published fixes
Epoch AI gave GLM 5.3 a deliberately low-tech setup and only information public at each release, and it built web pages that read local files from visitors on all six Firefox versions released between 17 September and 26 October. Once the environment was built, each attack cost about $1,000 in compute and the fastest took 12 hours, though the whole Firefox project ran over $100,000 in API spend. The newest versions also needed a separate unpublished bug, which Mozilla fixed in Firefox 159.0 on 27 October. A similar budget against Chrome did not succeed, and no provider safeguards were triggered.
safetysecuritySource: Epoch AI
Anthropic adds a rule steering Claude Code away from cloning specific games
Anthropic's Addy Osmani confirmed the company added two sentences to Claude Code's prompt on Opus 5.5 so it won't recreate someone else's specific game, citing the usage policy on IP infringement. Original games, genre games and games built from your own designs and assets are meant to keep working. The rule was spotted by a user, appears to have arrived around 27 October and can't be switched off, which is a reminder that vendor prompts can change your agent's behaviour without a version bump.
harnessespolicySource: Game Developer
Anthropic, Google, Microsoft and OpenAI back a shared opt-out from AI training
The four companies say they will honour a common signal, developed in an IETF working group, that lets a site opt out of generative AI training separately from search indexing. Microsoft says publishers can choose about training, grounding or both and still appear in Bing. Anthropic has already implemented it, while Google, Microsoft and OpenAI plan to do so in 2027.
policypublishingSource: Thurrott.com
Report: Anthropic in talks to lease a gas-powered data centre campus near Edmonton
The Globe and Mail, citing two unnamed sources, reports Anthropic is negotiating to be the tenant of a Beacon Data Centers campus in Sturgeon County, Alberta; both companies declined to comment. Regulatory filings show four data halls and a 920 MW gas-fired plant, which Alberta's Utilities Commission approved on 23 October with service due in late 2027. That would be about double the 420 MW planned for the Nova Scotia campus Anthropic leased earlier this month.
industrycomputeSource: The Globe and Mail
Claude Code 2.1.316 lets you set subagent effort and turn off ultrathink
The latest Claude Code release adds CLAUDE_CODE_SUBAGENT_EFFORT, so subagents can run at a set effort level instead of inheriting the session's, and CLAUDE_CODE_DISABLE_ULTRATHINK to turn that keyword off. Teams using the Claude apps gateway also get pooled group budgets, a single monthly budget shared by a group's members on top of each person's own limit.
toolsreleaseSource: GitHub (Anthropic)
Concept of the day
The patch gap
The patch gap is the time between a security fix becoming visible, for example in a public code change, and that fix actually reaching users' machines. During that window the fix itself points attackers at the flaw. Epoch's experiment shows open-weight models can now read those fixes and build working exploits cheaply, which shortens the safe window and pushes projects to ship updates faster.
Source: Epoch AI
Models released
| Model | Maker | Size | Notes |
|---|
Research
Context compaction quietly drops the rules agents are told to follow
Researchers tested five compaction strategies inside the OpenCode scaffold with three frontier models on PaperBench, ProgramBench and PostTrainBench. None beat default self-summarisation, but on ProgramBench the first summary already dropped about a third of the anti-cheating rules on average, and in real sessions instructions about how to work were dropped about six times more often than specs of what to build. Keep critical constraints in a file the agent rereads, not only in the chat.
researchagentsSource: arXiv
Some official low-bit Ollama models barely work at code
An audit of 305 code-capable GGUF tags in the official Ollama library found seven (2.3%) that solved at most six of 164 HumanEval+ tasks, all low-bit builds of models with four billion parameters or fewer. The files weren't corrupted; adding an importance matrix during quantisation lifted Qwen2.5-Coder-3B from zero to 76 to 113 tasks. The author releases a quantcheck tool and suggests screening small low-bit builds before trusting them.
researchlocalSource: 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
Five Agent Tricks, latest: Issue #5 (16 Nov 2026) · Subscribe free