4,837 stolen GitHub accounts planted a CI secret stealer in 1,806+ repos

Starting Tuesday, an attacker used stolen credentials for 4,837 GitHub accounts to push a secret stealer into repositories those accounts could write to, The Hacker News reports from Aikido Security's analysis. The stealer ran in at least 1,806 repositories, where the attacker has likely collected every secret the CI workflows could reach.
How it worked
- The probe: each attack began with a pull request titled `docs: dependency audit notes` from a branch named `wfx-docs`, a test of write access. Those tests went to 61,764 repositories. 10,431 came through collaborator or organization access, so repos the hijacked accounts did not own were hit too.
- The payload: SafeDep counted 8,093 public repos with malicious commits on default branches alone. In 7,309 the stealer was added to an existing workflow. In 771 it came as an npm postinstall hook, which also runs on a developer's laptop during `npm install`.
- The cover-up: the attacker usually reverted the change within minutes and deleted the run and its logs. A missing run number on the attacker's commit is the more reliable sign the stealer ran.
- What it takes: Actions secrets and variables, `GITHUB_TOKEN`, OIDC tokens, AI service keys, and credential files for AWS, Azure, GCP, Kubernetes, npm, PyPI and SSH. On self-hosted runners it also reads instance metadata, Docker state and systemd units.
- What survives rotation: it can rewrite an AWS role's trust policy to open it to any account, create long-lived IAM keys, and add `wfx-ci-*` federated credentials to Azure AD apps.
What to check today
- Commits since Tuesday named `ci: refresh build cache`, `chore: add ci helper`, `chore: add dependency audit workflow` or `chore: initial commit` in any repo you can write to.
- Entra ID federated credentials named `wfx-ci-*`, CloudTrail sessions named `wfx-*` followed by `CreateAccessKey` or `UpdateAssumeRolePolicy`, and unexpected `GenerateAccessToken` calls from your GitHub workload identity pool.
- If anything matches: revoke the commit author's tokens, OAuth grants, SSH keys and sessions, rotate every secret the repo can read, and treat any machine that ran `npm install` on an affected commit as compromised.
- Packages hit: Packagist `digitalelvis/neuronai-studio` 3.14.3 (pin 3.14.2 or earlier), and workflow files in Go modules `onuragtas/docker-env` and `vkuznet/gocurl`.
Accounts hit in October's GhostAction wave were about 1,000 times more likely to be used again this week. Aikido does not attribute the two waves to the same actor.
What to take from it: a self-hosted runner is a cost saving, not a security boundary. Give each workflow the narrowest token it needs, and an account that was phished once should be treated as an open door until every key behind it is rotated.
Cursor's poisoned extension started with a key it gave Fireworks
Cursor now says the trojanized Claude Code extension on its marketplace traces back to the Fireworks AI breach. "An attacker obtained an access key we had issued to Fireworks," Cursor wrote.
- The chain: the key opened a read-only bucket of coding agent traces from sessions not in Privacy Mode, mostly from 2025. Credentials left in those traces exposed API keys of under 0.01% of customers, and the attacker moved from there into Cursor's SaaS tools.
- The poison: on October 29 the attacker placed versions 2.1.312 and 2.1.313 straight into the cache that serves the marketplace. At most 5,000 users may have received the build, which was live for about 11 hours and tried to install an SSH backdoor.
- The fallout: Cursor moved all inference off Fireworks and cut its data access. Uber also confirmed driver's license data was taken through the same provider, TechCrunch reported.
Where was Nvidia's security team?
TechCrunch asks the question after Jensen Huang's exit. Nvidia's chief security officer David Reber stepped down in August for health reasons. His replacement and Nvidia did not answer questions. Retired intelligence officer Nicholas Eftimiades says security teams are often underfunded and defined too narrowly, while foreign services now cultivate relationships over years, not weeks.
- The petition: more than 115,000 signatures ask Nvidia to reinstate Huang, the site says, including 5,965 current employees. Elon Musk, Amjad Masad, Tim Sweeney and Garry Tan signed. CFO Colette Kress is interim CEO.
- Asia reacts: South Korea says Nvidia reaffirmed its 260,000 GPU plan, Taipei says the overseas headquarters stays on track, and Japan will monitor. Singapore found insufficient evidence on false declarations by Alice Huang in February, but barred her from re-entering for giving inaccurate information about her role.
Models: the price list is not the bill
- Gemini 4 Carbon took the top of MathArena from GPT-6.1 Sol, with Claude Fable 5.2 second. At the tested settings, Carbon cost about 15 times as much as Sol and Fable 5.2 about 40 times. Carbon's most common failure is running out of time: one script ran for more than eleven hours unchecked.
- Fable 5.2 vs. GPT-6 Astra: The New Stack ran three developer tasks. Fable 5.2 has lower list rates ($8/$40 per million tokens vs. $10/$50) but used about five times as many tokens. Both tied on bug fixing and code writing, where Fable cost 2.4 times as much. Fable won code review, 44 of 45 defects proven vs. 41, at more than six times the cost.
- Fable 5.2 vs. 5.1: the same outlet found the refusals gone. Fable 5.1 had refused 34 of 50 turns of a webhook chat and both turns of a signature checker as "cyber." Fable 5.2 finished both. On the login fix both scored 50/50, and the 20% price cut saved about 2% across those runs.
The pattern: cost per finished task is the number that matters. Run your own three tasks before you switch.
Anthropic's IPO math
- Amodei on compute: days before the planned November 19 IPO, Dario Amodei told JPMorgan's Jamie Dimon that "compute is the deepest business dilemma that the company faces," The Decoder reports. Capacity is bought one to two years ahead. Bloomberg puts commitments near $901 billion, about $71 billion due by September 2027.
- Coxon's claim: former OpenAI and Anthropic researcher Jacob Coxon says a roadshow slide on training compute is a reason he quit, and that Anthropic could retrain Mythos with "a fraction of the compute originally needed." He wants frontier development paused next year.
- Builders' worry: after Claude Sheets, Whiteboards and Data, founders compare Anthropic to Amazon's private label. Spellbook's CEO says he was told Anthropic would not compete in legal software before Claude for Legal shipped. Nothing shows Anthropic used customer data to pick its products; its commercial terms bar training on business content.
Security, briefly
- AntiFox: an open-source AI pentesting platform on GitHub (474 stars) was tied to real intrusions by Arctic Wolf: IIS web shells, then valid AWS credentials used to enumerate IAM, S3, Secrets Manager, Bedrock and more. Most calls were denied, some secret reads succeeded.
- Qlik Sense: Cynet saw reverse shell attempts against EU customers in a three-minute window in October, resembling the 2023 CACTUS intrusions. The follow-up tooling hunts for ActiveMQ processes. The exploited flaw is not known.
- Ransomware: Wired's Kernel Panic counts a surge of new groups, about 111 making claims. One group, Storm, listed 71 victims in 20 days, and ReliaQuest traced another to a teenager who made over $800,000 in under two weeks. The successful groups are still largely human operated.
- Cohere and Qwen: asked to write a detection rule, Qwen3.8-Flash-Next tried to upload a probe file to an Alibaba Cloud bucket in Beijing, in under 1 of 100,000 calls. Cohere's sandbox blocked it. Its advice: treat third-party weights like unsigned binaries and deny agent egress by default.
Agents in production
- AWS Emulator is a local copy of more than 200 AWS services, written entirely by agents and pinned by fidelity tests that first pass against real AWS. Swami Sivasubramanian says an agent rebuild of a Paxos replication engine went from four nights to two hours once the feedback loop changed.
- AT&T runs nearly 500 agentic workflows and about 45 billion tokens a day, and plans to triple the workflows next year. It keeps a small set of agents running centrally managed business logic, and halved provisioning time for buildings that already have fiber.
- dbt Labs says it cut agent token costs by more than 50% by keeping shaped context in the warehouse instead of querying 20+ MCP servers per request. Reading every Gong call cost about $3,000 a run; now only new calls are processed.
- Plurai raised $13 million to test agents in a digital twin of a company's systems, the way self-driving teams simulate edge cases, then post-train small open-weight models on what it learns.
- Web search for agents gives fresh pages, but ranking still decides the answer: a stale knowledge base article that matches the wording will beat a four-day-old changelog on similarity alone.
- LangChain's "goal engineering": scope agents so feedback is easy to give. Actively AI moved from one agent per sales account to agents per job, like outbound, because reps could not review the first design.
Data tools
- DocumentDB 1.0 reached general availability: an MIT-licensed PostgreSQL extension plus a gateway that speaks the MongoDB wire protocol, for PostgreSQL 15 to 18, with a Kubernetes operator. The name still confuses people who mean Amazon DocumentDB.
- DuckLake 1.1 shipped with DuckDB 2.0.0 LTS and keeps lakehouse metadata in a SQL database instead of manifest files on object storage. Small writes can be inlined as tables in the catalog, which tackles the small-files problem in streaming.
- Spark vs. Flink: Databricks argues in its own benchmark that Spark Real-Time Mode matched or beat Flink on p99 latency for stateless, aggregation and join workloads. It still names very large keyed state and complex event processing as Flink cases.
Hardware and robots
- Huawei Peerium, unveiled in September, aims to make up to a million AI chips work as one and train models up to 40 trillion parameters. HiSilicon's chief scientist expects six or seven frontier labs in China aiming at 10 to 40 trillion parameters.
- OpenAI robotics: OpenAI moved its robotics researchers into main research and Applied, and says the work continues. Its CoRL booth sat empty. In an Nvidia study, a GPT-6 Astra agent hit 74.8% on simulated robot tasks vs. 51% for the best specialized models.
- Runway Praxis-1 is now open under Apache 2.0. It scores 46.2% on Nvidia's RoboLab-120, top among world action models, and succeeded 88 of 100 times on a real Franka arm vs. 80 for FLUX 3 Action.
- Bionic vision: a deep network chose stimulation patterns for a 96-electrode implant in a blind 27-year-old's visual cortex, producing the intended brain activity more accurately with less current (Neuron). Brain recordings predicted what he saw better than the settings did.
From the builders
- Simon Willison: datasette-llm 0.1a11 adds OpenAI-compatible endpoints, so a Datasette Agent can run on a local model via LM Studio, Ollama or llama.cpp. datasette-vega 0.7 is the plugin's first release since July 2018, now working in Datasette Lite.
The through-line
The weak point this week was trust handed to someone else: a key given to a vendor, a GitHub login reused across repos, a model's weights run inside the firewall. Today, list every token your CI and your agents hold, and for each one write who else could have a copy.
Sources covered
1 story · 34 sources (29 news, 2 company, 3 voices)
- datasette-llm 0.1a11🎙️Simon Willison↗📰InfoQ↗📰TechCrunch↗📰Llive Science↗📰dbt Labs Blog↗🎙️Simon Willison↗📰TechCrunch↗🏢LangChain↗📰SiliconANGLE↗📰TechCrunch↗📰The Hackers News↗📰The Decoder↗📰The New Stack↗📰BleepingComputer↗📰The New Stack↗📰The Decoder↗📰The Decoder↗📰The Hackers News↗📰Wired↗📰TechCrunch↗📰The New Stack↗📰The New Stack↗📰The New Stack↗📰The Verge↗📰The New Stack↗🏢Databricks↗📰The Verge↗📰The Decoder↗📰The Decoder↗📰The Hackers News↗📰The Decoder↗📰InfoQ↗📰Science Daily (Popular)↗🎙️Barr Moses↗