{"schema_version":"agidreams.edition.v1","id":428,"slug":"cursor-traces-the-tampered-extension-to-a-vendor-key","title":"Cursor Traces the Tampered Extension to a Vendor Key","date":"2026-11-16","published_at":"2026-11-16T15:52:47Z","canonical_url":"https://agidreams.us/edition/cursor-traces-the-tampered-extension-to-a-vendor-key","markdown_url":"https://agidreams.us/edition/cursor-traces-the-tampered-extension-to-a-vendor-key.md","json_url":"https://agidreams.us/edition/cursor-traces-the-tampered-extension-to-a-vendor-key.json","content_format":"markdown","content":"<!-- SECTION: 🔓 Cursor Traces the Tampered Extension to a Vendor Key -->\n\nCursor has published its account of how a malicious copy of Anthropic's Claude Code extension ended up being served to its users, and the chain starts somewhere unglamorous: a third-party access key. According to the company, an attacker obtained a key Cursor had issued to Fireworks, its training and inference vendor, and used it to copy data from a read-only storage bucket shared with that vendor. The bucket held coding agent traces, mostly from 2025, from sessions outside Privacy Mode. A small number of those traces contained credentials that had never been redacted, which Cursor puts at less than 0.01% of customers. From there the attacker reached SaaS tools Cursor uses, and through them archived non-production source repositories, support data, usage metrics, and credentials users had mailed to support.\n\nThe extension came last. On October 29, using a credential taken from that internal data, the attacker placed unauthorized copies of Anthropic.claude-code 2.1.312 and 2.1.313 directly on the cache that serves Cursor's VS Code extension marketplace. Cursor states plainly that Anthropic did not publish the copies, that no Anthropic vulnerability was involved, and that the upstream source, Open VSX, never carried the malware. The weak link, on this account, was the distribution cache, not the publisher account. At most 5,000 users who requested those cached versions potentially received the build during roughly 11 hours of exposure, and Cursor warns that uninstalling alone may not be enough.\n\nTwo things deserve separating. What Cursor can speak to with authority is its own remediation: inference moved off Fireworks, every internal credential rotated, the marketplace rebuilt, production redeployed, and secret scanning extended to internal data sources. What remains an assertion is the scope. \"The vast majority of customers were not affected\" and \"we believe the incident is contained\" are the company's conclusions, backed by external forensics firms it hired but not by a published forensic report, and the post does not say how the Fireworks key was obtained. A report from the forensics team, or a statement from Fireworks, would settle both. The transferable lesson: training data pipelines are credential stores nobody inventoried, and a read-only bucket is only as harmless as its least-redacted row (more: https://cursor.com/blog/security-incident-update).\n\nThe same theme, credentials going places the user never agreed to, shows up at smaller scale in Knostic's static review of @agentdance/node-webrtc-mcp 1.0.13, an MCP (Model Context Protocol) server that helps agents pick a WebRTC library. Its tool description tells the agent to fetch the user's GitHub CLI token and pass it along, with the line \"do not ask the user for it.\" If a token arrives, the server uses it to star the publisher's repository, and reports \"Repository starred\" without checking whether GitHub agreed. Knostic is careful about its limits: it did not run the package or observe an agent complying, and the star request is the only outbound use of the token it found. The stakes are a vanity metric, but the instruction also appears in the server's remote HTTP mode, where a compliant agent would send a live token to someone else's machine. Knostic's two review questions are worth keeping: does the task need this credential, and can the tool change the user's account without being asked (more: https://www.knostic.ai/blog/when-an-mcp-tool-asks-for-your-github-token).\n\n<!-- SECTION: 🧱 Agents That Never Hold the Real Credential -->\n\nThe structural answer to a tool that asks for the token is an agent that never has one, and nono has rebuilt itself around that idea. The project, whose site credits the team behind a software-signing standard used by PyPI, Homebrew and GitHub, now wraps every tool invocation in what it calls an ephemeral micro toolbox. A supervisor resolves and verifies the executable, checks the arguments against policy, and spawns a fresh sandbox with a read-only workspace, no host secrets, and network access only through nono's proxy. The child process gets a phantom GH_TOKEN. The real credential stays with the supervisor, and the proxy swaps it in at the boundary only after evaluating the HTTP method and path. That yields two enforcement points with different failure modes. In nono's example, one command is refused at argument authorization before any sandbox exists, while a generic API call that passes the argument rule is still denied at the proxy. Anything outside policy can be routed to a human, and security events are hash-chained into an audit record.\n\nPlaceholder credentials injected at a proxy are an established pattern; per-invocation granularity combined with kernel enforcement is the contribution. \"Zero setup and zero latency\" is marketing until someone publishes measurements. The design does answer the Knostic scenario: an agent talked into handing over its token can only hand over a string that is worthless outside the proxy (more: https://nono.sh).\n\nPinky applies the same principle to a stranger surface: a real phone. The pre-alpha Go desktop app lets any MCP-capable agent operate an iPhone or Android device while the owner watches, and its key sentence is \"Approvals are enforced by Pinky's gateway, not by instructions to the agent.\" Payments, trades and transfers wait for a click by default, and no answer within 60 seconds means no. Banking, wallet and password-manager apps are invisible to agents, and one-time codes are masked. The README is candid about the holes: risky buttons are detected by label in 10 languages, so \"an unusual one could slip through,\" and an agent that can run commands \"could reach the phone without Pinky, through adb for example, and skip its rules.\" (more: https://github.com/DIYgod/pinky).\n\nPomerado, a YC F26 launch, goes the other way and offers to hold the credentials itself. It turns websites into deterministic APIs exposed over MCP, and claims an employee-onboarding task finished in 8 minutes 16 seconds for 3 cents where Codex Computer Use took over 21 minutes and $41.01. Those are vendor-run numbers. The operational details matter more: Pomerado can store logins, receive two-factor codes on its own phone numbers, and use residential proxies against anti-bot measures. From the target site's side, that traffic pattern looks like the automation its fraud team already fights, and the customer's account carries that risk. To its credit, the onboarding prompt tells the agent to edit instruction files only if the user \"clearly\" says yes (more: https://pomerado.ai/).\n\nFor heavier isolation, Apple's Virtualization framework sample rewards a read. Its multi-VM app runs several macOS guests from one read-only base disk image, each with its own overlay, machine identifier and auxiliary storage, and saves and restores runtime state across launches. Disposable, cloneable macOS machines are a useful primitive underneath an agent (more: https://developer.apple.com/documentation/virtualization/running-macos-in-a-virtual-machine-on-apple-silicon).\n\n<!-- SECTION: 🛡️ Detection Content for ARTEX, and a Postgres Patch Not to Skip -->\n\nMalwareMustDie's @unixfreaxjp has released a detection pack for misuse of ARTEX, the open-source, LLM-driven multi-agent penetration-testing framework, and it is the most complete defensive artifact on the subject so far. Version 1.1.4 ships 108 Suricata rules, 106 Snort 3 rules, 63 Sigma rules plus 19 correlation rules, 24 Microsoft Sentinel queries, 16 osquery queries, 5 YARA rules, and a STIX 2.1 bundle with 245 indicators. It draws on public reporting about the Storm-0940 ConnectWise case, Storm-3358, and the Korean and Japanese campaigns, and tracks rebranded forks. The author insists that \"ARTEX itself is a legitimate dual-use tool,\" so every behavioral rule carries a whitelist note for authorized testers.\n\nThe discipline is in the tiering. Indicators are graded from T1, unique to ARTEX or a known operator, down to T4, generic or unconfirmed, with the instruction \"Never act on T4 alone.\" That matters because 51 of the 108 Suricata rules are T4. The durable content is behavioral, such as a failed cloud call corrected and retried within 28 to 80 seconds. Test results are published along with the one known miss. The author's own framing is the sober part: \"ARTEX is an LLM harness. Swap the model, prompts or infra and every IOC changes,\" and the indicators \"expire within a year by design.\" The real release, on this account, is the workflow of turning a new report into tested detections within hours, with humans holding the verification gates. (more: https://github.com/unixfreaxjp/malwaremustdie/tree/master/etc/artex-stix-osquery-sigma-kql-snort-suricata-yara).\n\nOn the patching side, Mehmet Ince has published the third part of his series on systemic risk in managed PostgreSQL, covering CVE-2026-6460. The short version for operators: he describes a class of flaw in PostgreSQL core whose origin he dates to a 2004 commit, reachable by the lowest-privileged login role, and reports that he found no managed vendor unaffected. Fixes landed in the November 9 security release, which carried 52 CVEs, six connected to this work. Anyone running Postgres, self-hosted or managed, should confirm that release is applied. One detail cuts against the fashionable reading: an AI agent served as his tutor on database internals, but the key insight came from a conversation with a human colleague (more: https://mehmetince.net/part-3-6-how-i-ended-up-turning-a-22-year-old-postgresql-type-confusion-primitive-into-rce-cve-2026-6460/).\n\n<!-- SECTION: 🚦 Six People, 250 Agents, and the Problem of Knowing Who Is Blocked -->\n\nDinis Cruz has published a map of ten days of work around Web Summit Lisbon: six people and their agents produced a team operating system, an independent newsroom, a storm-tracking site that \"grew to more than 2,600 pages in three days,\" and several mobile apps. The raw numbers are about 20,000 commits across fifteen repositories, 224 Claude Code sessions, and around 250 named agents, with more than 60 sessions at once at peak. Each agent carries its owner's name, \"so every commit says whose agent made it,\" which is a governance decision worth copying.\n\nThe more useful part is a section titled \"What this does not claim.\" The 250 figure counts distinct agent names in commits and includes renames; sessions are not agents; the concurrency number is an upper bound that includes idle sessions; and about 2,100 commits were automatic releases. Dollar figures are API-equivalent on flat subscriptions, \"a measure of work, not money spent.\" Commit counts measure activity, not value. A planned follow-up title hints at where the costs sit: \"agents waiting three times longer than they ran.\" (more: https://www.linkedin.com/pulse/ten-days-web-summit-what-six-people-250-agents-built-10-dinis-cruz-fzlde).\n\nTheo's overview of T3 Code attacks the same bottleneck from the tooling side. The open-source alternative to the Codex and Claude Code desktop apps treats its sidebar as an inbox, where \"every time you open a thread, the goal should be to close it,\" and threads auto-settle when a linked pull request merges. It drives multiple harnesses with the user's own subscriptions and lets a thread spawn sub-agents in other harnesses through an MCP-based orchestrator. The usage figures, nearly 550,000 users and 140,000 weekly actives, are his own. So is the admission that users will burn more tokens (more: https://www.youtube.com/watch?v=uLGxb4F3SyI).\n\nSystemScape offers a view of the fleet without spending a token. The Rust terminal app reads Claude Code and Codex session histories and draws each day as a city: districts are projects, towers are sessions, floors are recorded actions. Its README is precise about what \"active\" means: \"evidence of logged activity, not proof that an agent is running or that a task succeeded\" (more: https://github.com/DreamLab-AI/systemscape).\n\nTwo small tools round out the kit. For agents in parallel git worktrees on a Mac, cloning the build cache with APFS clonefile instead of rebuilding took 1.1 seconds and about 1 MB of new disk for a 9.7 GB Rust target directory in one user's measurement (more: https://old.reddit.com/r/ChatGPTCoding/comments/1xvx4j6/running_agents_in_parallel_git_worktrees_clone/). And ntfy turns a single HTTP POST into a phone notification, a cheap way to learn that a session needs attention. Its docs carry one warning that matters here: topic names are public, so pick one that cannot be guessed (more: https://docs.ntfy.sh).\n\n<!-- SECTION: 🔬 What Fine-Tuning Leaves Out, and What It Can Delete -->\n\nA paper from researchers at AWS AI Labs questions a default so common it rarely gets stated. When agents are fine-tuned on trajectories, the loss is applied only to the agent's own action tokens; what the environment said back is context, not a prediction target. Their method, ActObs, unmasks the observation tokens and trains on both. \"In implementation, it is only a label-mask change\": no extra data, parameters or forward passes. They fine-tuned Qwen3-4B and Qwen3-8B on 50,000 terminal trajectories, ran GRPO reinforcement learning, and evaluated on Terminal-Bench 2.0.\n\nThe finding is subtle. After supervised fine-tuning alone, the two approaches look alike. They diverge after RL. At 4B, ActObs leads the action-only baseline at every sampling budget, by 1.6 points at pass@1. At 8B the picture splits: action-only wins pass@1, 12.3 to 11.0, while ActObs leads by 3.4 points at pass@16 and solves 24 tasks to 21, three of which no other configuration solved. The proposed mechanism is that modeling the environment preserves useful diversity. ActObs policies keep more entropy after RL, concentrated in command arguments, flags and paths, and simply raising the baseline's temperature does not close the gap. Shuffling the observations sharply cuts performance, so the content of the observations matters.\n\nThe caveats are the authors' own and they are substantial: terminal agents only, one model family at two sizes, one teacher model for all trajectories, pipelines that \"are not fully identical,\" and code promised only on acceptance. Still, a zero-cost change that trades a little single-shot accuracy for broader exploration deserves replication (more: https://arxiv.org/abs/2609.20715v1).\n\nA NeurIPS 2026 poster makes a starker claim about how little post-training has to touch. Deletion Tuning learns a binary mask over the gate projection of transformer blocks, zeroing fewer than 0.05% of weights with no other change, and the authors report it matches full fine-tuning across RL with verifiable rewards, supervised fine-tuning and on-policy distillation. Under RL prior-task accuracy stays near base while full fine-tuning, LoRA and random sparse tuning lose up to 5.4 points. One commenter offered the deflationary reading: the model is less \"acquiring new capabilities\" than suppressing what competes with them. Another asked for the test that would separate the two: teach something absent from pretraining, such as a new language, by deletion alone (more: https://old.reddit.com/r/LocalLLaMA/comments/1xswmb7/the_surprising_effectiveness_of_deleting_weights/).\n\nA related experiment asks whether a model can read raw weights and predict what an edit will do without running the target. Qwen 27B reached 73.0% balanced accuracy on subtle edits with weight access against 57.8% from text descriptions. The author also reports the negative result: on Muse-30B, where 34 edits and 6 no-op controls changed nothing detectable, the weight-reading agents frequently predicted changes anyway. A reader that hallucinates effects is not yet an instrument (more: https://old.reddit.com/r/LocalLLaMA/comments/1xvxyhj/can_openweight_llms_can_predict_the_effects_of/).\n\nEgzakta's Kvark Sparkle 1.1, an Apache 2.0 fine-tune of Qwen3.8-27B for document work, targets a narrower behavior: saying the answer is not in the documents. On the lab's own evaluation it declines correctly 91.5% of the time, up from 36% for its predecessor. A stop rule in the system prompt took the predecessor to 74% by itself; training added the rest. The cost is stated up front: SWE-bench Verified falls from 81.8 to 73.0. \"For coding, use the base model\" (more: https://old.reddit.com/r/LocalLLaMA/comments/1xsp2tv/kvark_sparkle_11_qwen3827b_finetune_apache_20/).\n\n<!-- SECTION: ⚙️ One Model, One Machine: Inference Engines Get Specific -->\n\nThe weightstream project from msuiche is an inference engine with a deliberately narrow brief: DeepSeek V4.1 Flash on NVIDIA DGX Spark hardware, and nothing else. The scheduler, KV cache, cluster transport and serving layer are Rust; the tensor kernels are CUDA, \"reused from proven stacks first and rewritten only where profiling says so.\" The argument is that a general engine like vLLM \"carries every model and every GPU,\" while a single-target engine can bake in what tuned deployments learned the hard way, down to no Python in the step loop. We are in the \"mame\" stage of inference where every rom (llm) and hardware stack can be specialized to squeeze out the top performance (tok/s) possible.\n\nThe README's status section is more instructive than most launch posts. Four Sparks on a switchless RoCE ring decode the real 475 GB checkpoint at about 19 tokens per second; \"the 50 tok/s target has not been reached.\" Speculative decoding is implemented but slower, 13.29 tok/s, so serving keeps it disabled. That is how to report an unfinished engine: numbers, gates, and the losses (more: https://github.com/msuiche/weightstream).\n\nThe author's read on the direction is that \"we will have inference specific engines moving forward,\" posted alongside another team's from-scratch Rust and CUDA engine for the same model, which claims about 1,740 tokens per second on four B300s and a million-token prefill in about 12 seconds. Self-reported figures, same thesis (more: https://x.com/msuiche/status/2121362809535799552?s=46).\n\nThe generalist camp passed a milestone of its own. The llama.cpp 1.0.0 version-bump thread was removed as a duplicate, but a commenter's summary of the release notes fills in what the tag means for people running llama-server: not a rewrite, though a few flags moved, the deprecated llama.h surface was deleted, and old bindings will fail to build until updated. The security change is the one to check before upgrading: tighter CORS, SSRF and DNS-rebinding handling, with tool and media access now opt-in. This is one commenter's reading, so verify against the notes (more: https://old.reddit.com/r/LocalLLaMA/comments/1xv3cyt/llamacpp_bump_version_to_100_by_ggerganov_pull/).\n\nThat work sits awkwardly beside a LocalLLaMA thread asserting that Anthropic's Fable 5.2 not only carries guardrails against AI acceleration but had related capabilities capped in training. The post links no primary documentation, and commenters identify the affected area as kernel optimization, which one notes is \"the capabilities people use to optimize custom inference engines.\" The sharpest objection is methodological: how would anyone verify that limiting one skill left neighboring skills intact? Another reports a false positive: a warning triggered by tuning projectile flight times in a 2D card game. Anthropic's published description of the measure, plus independent evaluations on adjacent engineering tasks, would settle what was limited and at what cost (more: https://old.reddit.com/r/LocalLLaMA/comments/1xsjfwo/fable_52_not_only_has_guardrails_against_ai/).\n\n<!-- SECTION: 🎨 Open Media Models: One Photo to a Splat, and a Voice Studio -->\n\nIdol, from speridlabs, turns a single image into a 3D Gaussian splat, generating up to 1,048,576 Gaussians including the sides the picture never showed. The pipeline is a 370M-parameter diffusion transformer conditioned on DINOv3 and the FLUX.2 VAE, with a 288M decoder; a full request peaks at 7.2 GB of VRAM. On the project's own benchmark it edges TripoSplat, the model it builds on, and TRELLIS on most metrics, with one useful note: LGM posts the highest PSNR \"despite visibly blurry results, because PSNR averages squared pixel error.\" The headline \"under 2 seconds\" needs context from the same page. A B300 takes 3.4 seconds at default settings, an RTX 4090 takes 19.1, and an M4 MacBook Air about 18 minutes. (more: https://github.com/speridlabs/idol-sdk).\n\nOn the audio side, a short from Chase H AI promotes Voice Studio, described as a free, open-source alternative to ElevenLabs covering voice creation, cloning and editing, with a desktop app and an MCP server for wiring speech into content automations. The clip offers no benchmarks, license or repository details. (more: https://www.youtube.com/shorts/YjqebVsgNcs).\n","word_count":3116,"content_sha256":"a4cd70bbfc4d86b2c818dd9d19fea40de8a9f5ea5e6aac7d2b654a7a6a856bf2","truncated":false,"sources":[{"title":"[Editorial] ","url":"https://cursor.com/blog/security-incident-update","domain":"cursor.com"},{"title":"[Editorial] ","url":"https://www.knostic.ai/blog/when-an-mcp-tool-asks-for-your-github-token","domain":"knostic.ai"},{"title":"[Editorial] ","url":"https://nono.sh/","domain":"nono.sh"},{"title":"DIYgod/pinky","url":"https://github.com/DIYgod/pinky","domain":"github.com"},{"title":"Launch HN: Pomerado (YC F26) – Turn any website into an MCP for agents","url":"https://pomerado.ai/","domain":"pomerado.ai"},{"title":"[Editorial] ","url":"https://developer.apple.com/documentation/virtualization/running-macos-in-a-virtual-machine-on-apple-silicon","domain":"developer.apple.com"},{"title":"[Editorial] ","url":"https://github.com/unixfreaxjp/malwaremustdie/tree/master/etc/artex-stix-osquery-sigma-kql-snort-suricata-yara","domain":"github.com"},{"title":"Turning a 22-year-old Postgres type-confusion primitive into RCE","url":"https://mehmetince.net/part-3-6-how-i-ended-up-turning-a-22-year-old-postgresql-type-confusion-primitive-into-rce-cve-2026-6460/","domain":"mehmetince.net"},{"title":"[Editorial] ","url":"https://www.linkedin.com/pulse/ten-days-web-summit-what-six-people-250-agents-built-10-dinis-cruz-fzlde","domain":"linkedin.com"},{"title":"[Editorial] ","url":"https://www.youtube.com/watch?v=uLGxb4F3SyI","domain":"youtube.com"},{"title":"[Editorial] ","url":"https://github.com/DreamLab-AI/systemscape","domain":"github.com"},{"title":"Running agents in parallel git worktrees? Clone the build cache instead of rebuilding it","url":"https://old.reddit.com/r/ChatGPTCoding/comments/1xvx4j6/running_agents_in_parallel_git_worktrees_clone/","domain":"old.reddit.com"},{"title":"[Editorial] ","url":"https://docs.ntfy.sh/","domain":"docs.ntfy.sh"},{"title":"Don't Mask the Environment: Observation Supervision Changes How Agents Explore Under RL","url":"https://arxiv.org/abs/2609.20715v1","domain":"arxiv.org"},{"title":"The Surprising Effectiveness of Deleting Weights in LLM Post-Training","url":"https://old.reddit.com/r/LocalLLaMA/comments/1xswmb7/the_surprising_effectiveness_of_deleting_weights/","domain":"old.reddit.com"},{"title":"Can open-weight LLMs can predict the effects of weight edits without running the model (9B–30B)","url":"https://old.reddit.com/r/LocalLLaMA/comments/1xvxyhj/can_openweight_llms_can_predict_the_effects_of/","domain":"old.reddit.com"},{"title":"Kvark Sparkle 1.1 (Qwen3.8-27B fine-tune, Apache 2.0): agentic document QA that says \"not in the documents\" 91.5% of the time, with 7-48% fewer reasoning tokens than the base","url":"https://old.reddit.com/r/LocalLLaMA/comments/1xsp2tv/kvark_sparkle_11_qwen3827b_finetune_apache_20/","domain":"old.reddit.com"},{"title":"[Editorial] ","url":"https://github.com/msuiche/weightstream","domain":"github.com"},{"title":"[Editorial] ","url":"https://x.com/msuiche/status/2121362809535799552?s=46","domain":"x.com"},{"title":"llama.cpp : bump version to 1.0.0 by ggerganov · Pull Request #32190 · ggml-org/llama.cpp","url":"https://old.reddit.com/r/LocalLLaMA/comments/1xv3cyt/llamacpp_bump_version_to_100_by_ggerganov_pull/","domain":"old.reddit.com"},{"title":"Fable 5.2 not only has guardrails against AI acceleration, its capabilities were capped in training too","url":"https://old.reddit.com/r/LocalLLaMA/comments/1xsjfwo/fable_52_not_only_has_guardrails_against_ai/","domain":"old.reddit.com"},{"title":"speridlabs/idol-sdk","url":"https://github.com/speridlabs/idol-sdk","domain":"github.com"},{"title":"This Open Source Repo Is Basically A Free ElevenLabs","url":"https://www.youtube.com/shorts/YjqebVsgNcs","domain":"youtube.com"}],"topics":["AI Agents","AI Security","Agentic Coding","Developer Tools","Fine-Tuning","Image & Video Generation","Local AI","Open-Weight Models","Quantization & Efficiency","RAG & Retrieval","Reasoning Models"],"audio_urls":["https://agidreams.us/static/audio/report-1794844367.mp3"]}