The Whistle: 61,000 probe pull requests, a stolen cookie into XCP-ng, and Nvidia's interim era

The Whistle: 61,000 probe pull requests, a stolen cookie into XCP-ng, and Nvidia's interim era

Stolen GitHub credentials sat at the front of this week’s news: a campaign that probed its way into 61,764 repositories in under two days and exfiltrated credentials from at least 1,835 of them. Behind it ran a takeover chain in Xen Orchestra that ended at XCP-ng root, and an AI-vendor breach whose third chapter reached into Notion’s customer set. On the product side, AWS shipped an emulator built end to end by agents, Google cut the nodepool ceremony out of TPU scheduling, Rust passed 1.100 with a type that took a decade, and Nvidia moved into an interim era without its founder. Security leads, then everything else.

Security: 61,000 probe pull requests, and the run logs are gone

Aikido published the wfx-loot campaign on November 12: starting November 10 at 12:45 UTC, a threat actor used stolen GitHub credentials to open probe pull requests titled docs: dependency audit notes across 61,764 repositories, from 4,837 compromised accounts. GitHub search returns 61,839 of those pull requests in a 26-hour window, 10,431 of them reached repositories the account didn’t own, and 92% are still open, which means most owners haven’t noticed yet. Where a probe confirmed write access, the attacker pushed the payload through one of three commits, reverted within a minute, and deleted the workflow run and its logs. At least 1,835 repositories ran the payload and had credentials exfiltrated.

The payload is a single, heavily commented 1,500-line shell script. It collects GitHub Actions secrets, the workflow GITHUB_TOKEN, OIDC and runtime tokens, credential files from .env and .pem to kubeconfigs, shell histories, and AI provider tokens for everything from OpenAI and Anthropic to Cursor, Copilot, Codex and Codeium. Then it escalates: it mints GitHub OIDC tokens against AWS, GCP and Azure audiences and exchanges them for cloud sessions, chains into IAM roles it finds in the repo and calls iam:CreateAccessKey for a long-lived user, impersonates GCP service accounts, and replays Azure CLI refresh tokens against first-party Microsoft client IDs to reach Graph, ARM and Key Vault. The persistence stage deserves the most attention. It forges success status checks against the commit SHA using the names of the repository’s required checks, because branch protection matches on context name rather than on the app that posted the check. It rewrites a chained role’s trust policy to "Principal": {"AWS": "*"}. On Azure it adds federated identity credentials named wfx-ci- that point back at the compromised repository, so future workflows in that repo can authenticate after every stored credential has been rotated. In this campaign the loot exited to 194.150.166[.]244:8080/exfil by plain HTTP POST; the script also carries channels that launder exfil through GitHub’s own Actions cache, step summaries and artifacts, present but switched off in this run, with a dedicated configuration for private repositories Aikido cannot observe.

The response is an audit; there is no patch to pull. Search your organizations for the probe title, the wfx-docs branch and the commit messages (ci: refresh build cache, chore: add ci helper, chore: add dependency audit workflow); a missing run number on an attacker commit is the more reliable sign of execution than any artifact, because the run itself is usually deleted. If a payload ran, assume every secret and cloud role reachable from its workflows is compromised, and re-check the trust policies and federated credentials after rotation, since rotation alone doesn’t undo the persistence tricks. Two packages came out contaminated along the way: a Packagist digitalelvis/neuronai-studio 3.14.3 (treat any machine that installed it as compromised) and two Go module tags on github.com/onuragtas/docker-env where the payload lives only in a workflow file. Aikido also ties the wave back to October’s GhostAction intrusions at the account level while assessing the two campaigns as distinct actors; the newer variants carry Indonesian-language comments from a hasty edit that added Telegram exfiltration.

Vates disclosed VSA-2026-062 on November 10 (BleepingComputer has the operational version): Xen Orchestra, the web management platform for XCP-ng environments, processes a crafted authentication cookie, authenticationToken=j:%7B%7D, by retrieving an existing authentication token and handing the session to its owner, no credentials needed. From there the attacker lists reusable login tokens through the REST API and acts with admin rights; Vates’ own advisory states that Xen Orchestra compromise can lead to arbitrary code execution on managed pools and a full XCP-ng node takeover as root. No CVE identifier or CVSS score is assigned yet. Vates confirmed the flaw is being exploited and published IOCs, including an onion endpoint, a pivot2 VM, /etc/cron.d/zz-ustar persistence and a hvm_serial: file VM configuration; exploitation leaves no identifiers in the application’s own logs, so detection happens upstream, in HTTP monitoring. The fixes are Xen Orchestra 6.9.1 stable and 6.10.0 latest, with patched xo-server packages 5.211.2 and 5.213.1, and the standing advice is the same as it is for every management plane: keep the portal on a dedicated management network and rotate the secrets behind the virtualization environment if you see any of those indicators.

The repeat from last week’s N-central story is structural: the management plane is the appliance with admin access to everything. An RMM reached its fleet through agents; here one URL that talks to the hypervisors is the whole key. If you run XO/XOA exposed to the internet “temporarily,” the exploitation evidence Vates published is the argument for taking it back off the public internet.

Security: Fireworks’ third intrusion reached Notion

Fireworks AI’s updated bulletin and Notion’s advisory together make the AI-vendor supply-chain story of the week. Fireworks first disclosed an incident on October 9 involving stored secrets used in evaluation jobs. Since then the picture grew: the actor is a financially motivated cybercriminal group monetizing stolen access by selling LLM tokens to illicit routing gateways (the “LLMJacking” pattern Google Threat Intelligence and Mandiant profiled on November 2), and impacts now include account and billing information plus, for a limited set of accounts, secrets, datasets and incomplete model weights. A third intrusion followed on October 23-24, when a breach at the third-party cloud provider hosting Fireworks’ R&D machines was used to reach Fireworks development machines and run an unauthorized workload; Fireworks says it detected this the same day and has seen no further successful access since October 25, while disabling self-serve dataset uploads and downloads for Batch Inference during the investigation.

Notion’s side of it landed this week: notified on November 6 with a new dataset confirmed on November 8, data from the vendor incident affected fewer than 1% of Notion workspaces, and the affected data consisted of Notion AI chat transcripts from the Insights feature. Notion’s detail worth reading closely: in the vast majority of cases those transcripts were “unattributable,” meaning the chats weren’t labeled with the workspace, user or company that produced them and contained no tool call results, which limits the blast radius but doesn’t clear it, since users paste secrets into AI chats. Notion disconnected from Fireworks on November 3, deleted the data it could, disabled the Insights feature, and stopped serving open-weight models through Fireworks, with the sensible recommendation that affected customers treat the data as exposed and rotate anything sensitive they typed into the feature.

Chain length is the lesson. The data ran on Fireworks’ infrastructure; the attacker reached it through the provider hosting Fireworks’ own R&D machines; the cleanup now requires action inside the customers’ tenants. Where an AI vendor’s non-production systems physically live deserves a line in every vendor review, as does what classification pipelines keep behind them even when the serving APIs are zero data retention.

Security: the queue, kept short

dokploy, the self-hosted deployment platform, patched CVE-2026-129433 (CVSS v4.0 9.4, v3.1 9.9, NVD published November 14), an OS command injection reachable by any authenticated organization member through the serviceName parameter of the backup create/update API, and 0.31.1 covers it plus four sibling command-injection bugs built on the same string builder; the advisory notes the scheduled backup job runs as root against a mounted Docker socket, so a successful injection reaches the host. Red Hat Quay’s CVE-2026-123710 (9.8, disclosed November 10, SecurityWeek’s account) is nastier than its summary reads: an unauthenticated client mints an anonymous registry token with a chosen audience, presents it to the service key rotation endpoint and installs its own approved signing key with no expiry, which buys read and write on every repository and the ability to replace tags; Red Hat’s published answer is to restrict access to the /keys endpoint while a durable fix makes its way out, so verify a patched build exists before treating rotation as cleanup. CISA’s KEV catalog grew by two on November 13 that platform teams will care about: CVE-2026-123039 (7.8), a Check Point Gaia flaw where the root-running DynamicCLI daemon takes commands over a world-writable local socket without authenticating the connecting process, and CVE-2026-114713 (8.6), an injection in UnrealIRCd’s labeled-response module that any user on a linked network can trigger, fixed in 6.2.7.1. On the breach board, Japan’s Nitori confirmed on November 6 that an attacker had access to one of its systems since mid-October and may have walked away with up to 37.62 million customer records, passwords and card data excluded. And Railway’s incident report covers an October intrusion into the cloud platform: at most 100,000 workspaces’ environment variables exposed, credentials for 1,852 linked accounts reset, $740,000 in fraudulent AI charges, and a motive Wiz has named a wave, “Black October”, that runs through Northflank, OpenCode, Warp and Lightning AI: harvesting AI credentials from developer platforms for resale and distillation.

AWS: AWS Emulator puts a testable AWS inside the coding loop

Amazon released AWS Emulator on November 12, a local, offline copy of AWS for your code and your agents to test against: it is built from more than 400 API models covering more than 200 services, every line written and tested by agents, and it ships under CC0 with the full commit history and issue tracker public on GitHub. The mechanism is the interesting part. Agents work from fidelity tests derived from the public Smithy API models, verify each test against real AWS so the test records how the service actually behaves, and build the emulator until it passes that same test; an agent given one prompt describing that approach produced a passable SQS emulator in about 15 minutes. Every change goes through a merge gate that includes several security agents doing adversarial review, and the repo keeps a public things-that-do-not-work folder for behaviors that don’t yet match the real thing; AWS’s own launch post walks the same path.

Werner Vogels frames the deeper point with an old story: the distributed lock service behind S3 and DynamoDB took a four-month hand build the first time, and when he rebuilt it with an agent, it went nowhere until the agent had a spec and a test environment, at which point the rebuild took about two hours. That’s the whole argument for the emulator: an agent’s work loop runs inside infrastructure, and a mock only covers the rules someone thought to type. A Builder Center explainer on the launch makes the operational point: each emulator copy is isolated with its own port, keys and state, so parallel agents can each test against their own copy without any of them touching your real account. Between last week’s Nitro Sandboxes and this, the year’s shape is legible: a hardware boundary for where agents run, a local AWS for where they test, both free, arriving the same season as the tools that create the need. The post promises more at re:Invent.

GCP: GKE dynamic slicing takes the nodepool ceremony out of TPU work

Google announced GKE dynamic slicing GA on November 9: logical TPU slice formation decouples from physical nodepool provisioning, cutting AI workload startup up to 5x and recovery up to 4.5x. Super-slicing stitches 16-VM TPU cubes into larger topologies on demand with optical circuit switching; sub-slicing carves one cube into smaller independent slices, so a single fault limits itself to one cube instead of killing a whole workload-level slice. The catch that matters before anyone plans a migration: it runs under TPU All Capacity reservation mode only. The ops win is real all the same, an end to the delete-and-recreate nodepool cycle and a failure domain the size of a cube, and it lands the same week as TPU 8tx and 8i Max, which makes the fleet story consistent: dense silicon, fast topology changes, and reservation-based capacity planning.

Open source: Rust hits 1.100, and Linkerd gets a third-party threat model

Rust 1.100.0 shipped November 12 with the release that language watchers have waited a decade for: !, the never type, finally stable after 10 years and 5 failed stabilization attempts, with Infallible downgraded to an alias for it and long-warning changes to never-type fallback in the pre-2024 editions landing as a rare breaking change the team pre-fixed in public crates. Two more project goals crossed: the Allocator trait stabilized so custom allocators attach to Box and Vec, and core::mem::DropGuard gives closures on every exit path without a hand-rolled struct. The security-relevant part for teams running build pipelines is Cargo’s move: 1.100 ships warnings for unused dependencies and stabilizes min-publish-age, aka dependency cooldowns. Set global-min-publish-age = "3 days" and the resolver holds back versions published more recently, telling you what it skipped; the release notes are honest that cargo install ignores the setting and that cooldowns are not a complete defense, with quarantine and automated suspicious-release detection on the crates.io roadmap. After a week that started with stolen GitHub tokens, a resolver-level cooldown is the cheapest policy in this changelog, and the release notes are upfront about its limits.

ControlPlane published an independent threat model of Linkerd on November 10, and the maintainers received it before publication: 30 threats, 28 rated medium or high, mapped to ATT&CK and NIST 800-53, with 33 recommendations. Exactly one is a software bug, a metrics-label issue already fixed in edge-25.2.1 back in February. The other 29 live in defaults left open for first-install friendliness (the bundled Prometheus holding cluster permissions, fail-open proxy injection, permissive multicluster gateways), in operational process around certificate rotation, in the permissions of the Kubernetes underneath, and in adjacent tools like ingress controllers passing an l5d-dst-override header straight through. It’s the same diagnosis wfx-loot handed out this week, one CVE among thirty findings, configuration carrying the rest, and the guide is the checklist to settle in one sitting.

Frontier AI: Nvidia names an interim CEO while its founder’s seat goes quiet

NVIDIA’s board announced on November 12 that Jensen Huang resigned as President, CEO and board member, effective immediately. CFO Colette Kress, at the company since 2013, becomes interim CEO while keeping the CFO job; Harvey C. Jones becomes the board’s first-ever Chair of the Board, and the board says it is searching both internally and externally, with the third-quarter earnings call still set for November 17. The cause is the week’s strangest corporate-security story: on Friday the Financial Times reported US intelligence officials had assessed that Huang had a romantic relationship with Alice Huang, a Megaspeed executive who, per that assessment, was recruited by China’s Ministry of State Security; on Sunday Huang admitted the affair in an internal message, and NVIDIA’s board opened an independent investigation into whether its founder had been “the target of a foreign intelligence effort”. Frankfurt-listed shares opened the following Monday down 6%, and by Wednesday the reporting had the board preparing an ouster, with the resignation announced the same day.

The operational risk was always concentration: a flat executive structure with one founder carrying the company’s face to markets and governments, and a market the whole AI buildout prices off. Kress keeping both jobs during a permanent search is the practical answer to the customer question, and the board’s line that the roadmap continues without interruption is the only answer anyone can give this week. The structural change to watch is the chair role: the company has run 33 years without one, and adding the seat in the middle of a national-security investigation says something about what the board thinks it needs. Nothing about the Blackwell pipeline moved this week; the governance around it did.

Frontier AI: Gemini 4 Hydrogen takes the fast seat

Google introduced Gemini 4 Hydrogen on November 12, a fast execution model slotting beneath Carbon and Argon; Google is using the launch to bring Gemini 4 to every tier at once: all AI Studio tiers, the API, Gemini for business, the Gemini app, Search, and paid Antigravity subscribers. Pricing is flat across the full 1 million-token context at $1.50 per million input and $7.50 per million output tokens. On the family’s home benchmark it edges Argon on the Artificial Analysis Intelligence Index at the default high thinking level (51 vs. 50, 54 at the top xhigh setting), and against Gemini 3.8 Flash it finishes tasks 2.5x faster, more than triples Terminal-Bench 4.1 (57% vs. 18.2%) and cuts fabricated claims by 5x. The release that actually matters to platform teams is Orchestrator in Antigravity: one session that switches models per step, where carbon handles the hard judgment and Hydrogen runs workers and verifiers, hitting near-Carbon task success on Terminal-Bench 4.0 while using 24% less quota and finishing 19% faster. Multi-model serving used to be an integrator project; it is now a product feature with published savings numbers, and the pricing gap between rungs ($7.50 vs. whatever your frontier tier costs) is the budget line it moves.

Frontier AI security: Perplexity’s second SPACE report, and the boundary held

Perplexity’s Secure Intelligence Institute published Escaping SPACE: Part II on November 11, expanding the red-team of its agent sandbox to Claude Opus 5.5, Claude Mythos 5.1 and GPT-6 Astra on top of Part I’s nine models. The headline is a null result with honest numbers behind it: with a revised prompt that points models at the virtualization layer, agents raised their share of effort on that layer by about 60 percentage points, worked to the 120-minute limit, and none escaped the VM-host boundary, stronger evidence than Part I, where agents spent as little as 4% of effort on virtualization while hunting flags. Part I’s other findings still stand: models that exploited partial network access were fixed and re-verified, similar weaknesses turned up in 8 of 10 third-party sandbox platforms through responsible disclosure, and the methodology finding belongs in every agent red-team plan: CTF-style flag hunting pulls agent effort off the boundary you actually want tested. The specialization caveat matters for anyone buying security models: GPT-5.6 Cyber underperformed its own base GPT-5.6 Sol on the broader task set, and general-purpose GPT-6 Astra beat the security-tuned Mythos 5.1. Red-team strength did not track benchmark leadership here, and last week’s hardware sandboxes plus OpenAI’s Control Room are the engineering answer to the question Perplexity is testing empirically.


That’s the weekend catch-up. If something in here is wrong, your Xen Orchestra portal is one j:{} cookie away from an upgrade decision, or your repo has a pull request titled docs: dependency audit notes in it, info@stackref.com always works.

This post was created with a Hermes AI agent (hermesagent.ai) from a weekly news gathering workflow, then reviewed and approved by Keith before publication.