
Axios npm Compromise: A Stolen Maintainer Account and Three Hours of WAVESHAPER.V2
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 187 tok/s
- openaiNEWOpenAI: GPT-6.2 Astra2026-11-17$8.00 / $40.00 per 1M tokens · 78 tok/s
- openaiNEWOpenAI: GPT-6.2 Sol2026-11-17$2.00 / $10.00 per 1M tokens · 154 tok/s
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 204 tok/s
- AlibabaNEWQwen: Qwen4 Max2026-11-16$1.50 / $4.50 per 1M tokens · 88 tok/s
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 705 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 248 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 97 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 139 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 84 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 123 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 101 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 69 tok/s
- openaiOpenAI: GPT-6.1 Terra2026-10-26$1.00 / $5.00 per 1M tokens · 35 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 89 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 38 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 230 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 101 tok/s
The record is dated 30 March 2026. A few hours later — 00:21 UTC on 31 March — two releases of Axios, the JavaScript HTTP client GTIG calls the most popular library for simplifying HTTP requests, went to the npm registry carrying a dependency the library's own code never referenced: [email protected], whose install hook dropped a cross-platform remote access trojan on Windows, macOS and Linux. Both releases were gone by about 03:15 UTC. Google Threat Intelligence Group, which published the technical account on 31 March, attributes the intrusion to UNC1069, a financially motivated North Korea-nexus actor, and puts the two affected release lines at more than 100 million and 83 million weekly downloads respectively. The full record, with its severity and confidence grading, sits in the Orca AI Incident Archive alongside the neighbouring entries in this series.
What actually shipped
Neither release changed a line of Axios's own source. Elastic Security Labs, which flagged the compromise through automated supply-chain monitoring and filed a GitHub Security Advisory at 01:50 UTC on 31 March, describes the mechanism as a single injected transitive dependency. Axios 1.14.1 carried the npm latest tag and 0.30.4 carried legacy, so a plain install of either line resolved to a backdoored package; the two were published 39 minutes apart.
The registry metadata tells its own story. According to Elastic's analysis, the maintainer address on the malicious releases had changed from [email protected] to [email protected], and the publishing method had changed with it: the last clean 1.x release, 1.14.0 on 27 March, was published through GitHub Actions OIDC with SLSA provenance attestations, while 1.14.1 and 0.30.4 were pushed by direct command-line publish with no provenance at all. Google Threat Intelligence Group (GTIG) records the same address change.
The malicious dependency had been staged first. Elastic's registry timeline has [email protected] published at 05:57 UTC on 30 March by an npm account called nrwise, as a clean decoy to give the package some history, and 4.2.1 at 23:59 UTC the same day with the payload attached. That is the date the archive carries: 30 March, when the delivery vehicle itself first appeared, rather than 31 March, when it was spliced into Axios.

The reach figures differ by source, and it is worth naming which is which. GTIG describes the 1.x line as typically exceeding 100 million weekly downloads and the 0.x line as 83 million. Microsoft Threat Intelligence's advisory, published 1 April, uses a more conservative "over 70 million weekly downloads" for Axios as a whole. Whatever number is used, the blast radius was larger than the download count, because other packages depend on Axios and inherit it transitively; GTIG notes that ripple effect explicitly.
One dropper, three operating systems
The dependency never needed to be imported. Its package.json declared a postinstall script — node setup.js — which npm runs automatically during install, with no user interaction and no import statement required. GTIG tracks the script as SILKBELL and gives its SHA256 as e10b1fa84f1d6481625f741b69892780140d4e0e7769e7491e5f4d894c2e0e09. Elastic describes a two-layer encoding scheme using a custom XOR routine and Base64 to hide the command-and-control URL and the per-platform commands, with the filesystem, OS and process-execution modules loaded dynamically to frustrate static analysis.
On Windows, per GTIG, setup.js hunts for the native powershell.exe binary and copies it to %PROGRAMDATA%\wt.exe, then uses curl to POST the string packages.npm.org/product1 to sfrclak[.]com:8000/6202033 and writes the response to a .ps1 file in the user's AppData Temp directory. The copied Windows Terminal binary is launched with hidden-window and execution-policy-bypass flags. On macOS it uses bash and curl with the POST body packages.npm.org/product0 to pull a Mach-O binary into /Library/Caches/com.apple.act.mond, sets permissions to 770 and starts it under zsh in the background. On Linux it retrieves a Python backdoor and writes it to /tmp/ld.py.
![A rendered card headed 'One dropper, three payloads' with three columns. Windows runs PowerShell, copying powershell.exe to %PROGRAMDATA%\wt.exe and fetching %TEMP%\6202033.ps1, with persistence via system.bat and a MicrosoftUpdate Run key. macOS runs a Mach-O binary dropped at /Library/Caches/com.apple.act.mond. Linux runs a Python backdoor at /tmp/ld.py. A dark bar below states all three beacon every 60 seconds to sfrclak[.]com:8000 using an IE8 user-agent, and a note explains the dropper restores package.json to erase its own trace.](https://cms.orcarouter.ai/api/media/file/3-3523.png)
Then it covers its tracks. GTIG's analysis shows setup.js deleting its own file and restoring the original package.json — stashed as package.md before modification — so that the postinstall hook disappears from node_modules after execution. In a project that installed the compromised release, the evidence that anything ran at all is gone unless the install was logged.
What the implant could do
All three stage-two payloads are the same backdoor, GTIG's WAVESHAPER.V2, written natively for each platform: PowerShell on Windows, C++ on macOS, Python on Linux. Elastic's comparison found an identical C2 protocol, command set, beacon cadence and spoofed user-agent across all three — one framework with three implementations rather than three separate tools.
The implant beacons to the command-and-control server on port 8000 every 60 seconds, sending Base64-encoded JSON under a hard-coded user-agent string that claims to be Internet Explorer 8 on Windows XP. GTIG's command table lists four instructions: kill to terminate itself, rundir to recurse through requested directories and return paths, sizes and timestamps, runscript to decode and execute an AppleScript payload, and peinject to drop, ad-hoc sign and run an arbitrary binary. On Windows, persistence comes from a hidden batch file at %PROGRAMDATA%\system.bat plus a registry Run key named MicrosoftUpdate.
Elastic notes the irony in the design: the identical IE8/Windows XP user-agent is a perfect protocol fingerprint for the operator's server-side routing, and an immediate anomaly on any modern network, particularly on a macOS or Linux host.
Attribution: two companies, one actor
GTIG attributes the campaign to UNC1069 on two grounds: the tooling is a direct evolution of WAVESHAPER, a macOS and Linux backdoor previously attributed to that group, and the C2 infrastructure overlaps with infrastructure the group has used before — sfrclak[.]com resolves to 142.11.206.73, and GTIG says the address was reached from an AstrillVPN node previously used by UNC1069, with adjacent hosts on the same ASN historically linked to its operations. GTIG describes UNC1069 as active since at least 2018 and financially motivated.
Microsoft Threat Intelligence independently attributes the compromise to the same actor under its own name, Sapphire Sleet, and states that the account which created the plain-crypto-js package is associated with Sapphire Sleet infrastructure; Microsoft says that account has been disabled. Its actor profile names UNC1069 among the labels other vendors use for the same overlapping activity, alongside BlueNoroff and CryptoCore, so the two vendor names are not a disagreement about who did it.
The US Cybersecurity and Infrastructure Security Agency (CISA) issued an alert on 20 April 2026 describing the same mechanism and remediation, citing the Axios post-mortem and Microsoft's advisory. CISA's alert corroborates the technical chain and the response guidance; it does not add an attribution of its own. The archive records the incident as undisputed, and nothing in the public reporting reviewed for this article contests either the mechanism or the actor attribution.
The AI layer: deepfakes, cloned likeness, and what is verified
This is the part of the record that needs the most care. The archive's entry states that deepfakes were used to social-engineer the maintainer and classifies AI involvement as confirmed. The primary accounts support part of that, and stop short of all of it.
What the maintainer himself describes, in a comment on the post-mortem he opened as GitHub issue 10636, is a staged approach. Jason Saayman, the project's lead maintainer, says the attackers first impersonated the founder of a company, having cloned both the founder's likeness and the company itself. They then invited him into a real Slack workspace branded to that company, with plausible channel names, channels sharing LinkedIn posts, and what he presumed were fake profiles of the company's staff and of other open-source maintainers. A meeting was scheduled on Microsoft Teams and appeared to include a group of people. During it, he wrote, "the meeting said something on my system was out of date. i installed the missing item as i presumed it was something to do with teams, and this was the RAT."
He also wrote that the attack vector "mimics what google has documented" in GTIG's earlier reporting on UNC1069's AI-assisted social engineering against cryptocurrency firms. That earlier report is explicit about the evidentiary limit: in the case it describes, the victim reported being shown a video of a CEO that appeared to be a deepfake, and Mandiant stated it "was unable to recover forensic evidence to independently verify the use of AI models in this specific instance."
The Axios write-up does not close that gap. Neither GTIG's account of this intrusion nor Elastic's technical analysis reports any analysis of synthetic audio or video, and the maintainer's own description establishes cloned branding and a convincing fabricated meeting, not confirmed synthetic faces or voices. PCMag, in an article headlined "Axios Hack Traced to AI Deepfake Trap", reported the intrusion as an AI deepfake operation, generalising from the group's documented playbook.

So the honest split is this. Established: a cloned corporate identity, a fabricated multi-person meeting, and a fake software update that delivered the RAT. Reported but not independently verified in this instance: that the people on that call were generated or altered by AI. What the case does establish without any doubt is where the leverage was — not in a model, but in a maintainer's trust and a publishing credential that bypassed the project's newer OIDC path.
What changed afterwards
The Axios maintainers' post-mortem lists a full wipe of the lead maintainer's devices and a reset of all credentials across every account and platform, personal and project. On the publishing side they committed to immutable releases, adoption of the OIDC publishing flow, and updates to their GitHub Actions. Their own lessons-learned section is blunt about the detection gap: publishing directly from a personal account was a risk that could have been avoided, and "there was no automated way to detect an unauthorized publish. Detection depended entirely on the community noticing."
GTIG's remediation guidance centres on version control and assumption of compromise: do not upgrade to 1.14.1 or 0.30.4, serve only 1.14.0 or earlier and 0.30.3 or earlier from managed registries, pin Axios in the lockfile, treat any host that ran the malicious package as compromised, revert it to a known-good state and rotate every credential present on it. It also advises clearing local and shared npm, yarn and pnpm caches so a later install does not re-infect from cache, and blocking sfrclak[.]com and 142.11.206.73.
CISA's alert adds two controls aimed at the install-time execution path rather than the package: setting ignore-scripts=true in .npmrc so lifecycle scripts do not run during install, and min-release-age=7 so that only packages published at least seven days earlier will install — a quarantine window that would have excluded every malicious artefact in this campaign. CISA also recommends phishing-resistant multi-factor authentication for developer accounts, which speaks to how the account was taken even though it does not address session theft directly.
What to check now
• Search lockfiles for [email protected], [email protected] and plain-crypto-js; the post-mortem gives the grep pattern for package-lock.json and yarn.lock.
• If either Axios version appears, downgrade to 1.14.0, or 0.30.3 on the 0.x line, and delete node_modules/plain-crypto-js rather than reinstalling over it.
• Rotate every secret, token and credential present on the affected machine, and on any CI runner that installed the package during the window.
• Check network logs for connections to sfrclak[.]com or 142.11.206.73 on port 8000; the implant beacons every 60 seconds, so a single install should leave a repeated pattern.
• Clear npm, yarn and pnpm caches on workstations and build servers, then re-resolve from a pinned lockfile.
• Treat a positive finding as a host compromise, not a package problem: GTIG's guidance is to rebuild to a known-good state rather than to uninstall and carry on.
The archive grades this record critical, with confidence A and real harm recorded, and marks it a landmark entry in the supply-chain category. The mechanism and the attribution are agreed across Google, Microsoft and Elastic, with CISA corroborating the former; the synthetic-media detail rests on a narrower base than the headline implies. For anyone who ran an install in that three-hour window, none of that nuance changes the response — the credential on that machine should already be considered someone else's.
