Skip to content

Contagious Interview SVG-steganography OtterCookie campaign

Summary

Elastic Security Labs reported on July 18, 2026 a new Contagious Interview-aligned developer-targeting campaign, tracked as REF9403, after DPRK-aligned operators targeted Elastic's community Slack with a fake job posting and coding-challenge project.

The durable defender point: the repositories looked like functional ecommerce / coding-test projects, but hid payload fragments inside HTML comments embedded in SVG flag images. Running the project caused a JavaScript loader to reassemble and execute the payload, leading to a four-stage chain aligned with OTTERCOOKIE: browser and wallet theft, file theft, Socket.IO remote access, and clipboard theft.

Tags

Reported chain

  1. The actor approached developers with a fake job posting and coding challenge / take-home assignment.
  2. The supplied project was functional enough to appear legitimate, reducing suspicion during review or execution.
  3. Payload chunks were Base64-encoded inside HTML comments across country-flag SVG files under an assets/flags directory, including normal-looking files such as AE.svg and AF.svg.
  4. A repository JavaScript file named serverValidation.js read all .svg files in sorted order, extracted the comment bodies, concatenated the fragments, decoded them with a custom Check() function, and executed the result with eval().
  5. Elastic described a four-stage payload aligned with OTTERCOOKIE:
  6. browser credential and cryptocurrency-wallet extension theft;
  7. local file theft;
  8. a Socket.IO-based remote access trojan capable of shell-command execution and follow-on payload delivery;
  9. clipboard theft.
  10. The browser/wallet module set its process title to npm-cache, enumerated Chrome-family browser profiles across Windows, macOS, and Linux, and collected Login Data, Web Data, and wallet-extension Local Extension Settings stores.
  11. Elastic reported multipart HTTP POST exfiltration to ldb.rightwidth[.]dev endpoints including /upload and /cldbs, with User-Agent axios/1.18.1.

Why this matters

  • Developer execution is the compromise boundary: the risky action is running an apparently benign project, not installing a package from a registry.
  • Static review can miss distributed payloads: malicious content is split across image files that still render as normal SVG assets.
  • The target set overlaps supply-chain blast radius: browser sessions, wallet extensions, source-control credentials, cloud sessions, package-registry access, and AI/developer-tool context can all be reachable from developer endpoints.
  • The campaign tracks current developer tooling: Elastic noted exclusions and file-scoping logic that show awareness of modern development environments and AI coding tooling directories such as .claude, .cursor, .gemini, .windsurf, .pearai, and .llama.

Defender heuristics

  • Treat unsolicited coding tests and take-home assignments as executable content. Review them in isolated environments with no logged-in browser, wallet, source-control, package-registry, cloud, or AI-assistant credentials.
  • Search repositories for SVG files containing long HTML comments or Base64-like blobs, especially under assets/flags or other asset directories where payload fragments may hide in otherwise normal media files.
  • Inspect project scripts for JavaScript that reads many asset files, sorts filenames, extracts comments with regexes such as <!-- ... -->, concatenates data, uses custom Base64 decoders, or calls eval() during server startup or validation routines.
  • Hunt on developer endpoints for node/project processes renamed or presented as npm-cache, unexpected Socket.IO connections, suspicious axios/1.18.1 POSTs, and requests to /upload, /cldbs, /api/service/makelog, or /api/service/process/<uid> on unfamiliar infrastructure.
  • Monitor for browser credential-store access (Login Data, Web Data) and wallet-extension LevelDB reads from project-launched Node processes.
  • If a coding-test project was run, isolate the endpoint before rotating credentials; then revoke browser sessions, source-control tokens, package-registry tokens, cloud/API keys, wallet secrets, and AI-tool credentials that were reachable from the host.

September 21 follow-up: the joint multi-agency advisory puts numbers on the campaign, 30,000+ devices, 100+ countries, 7,000+ wallets, $10.71M, and formally fuses WaterPlum IT-worker ops with Contagious Interview

The Hacker News (Sep 21, 2026) covered a joint advisory from Japan, the U.S., Australia, and Germany (canonical advisory URL not retrievable at capture, THN's article links only its secondary sources; backfill from CISA's joint-advisory index next sweep) quantifying the campaign this page documents fragments of: at least 30,000 compromised devices in 100+ countries, funds or credentials siphoned from 7,000+ cryptocurrency wallets, at least $10.71M plundered. Targets: web designers, engineers, crypto/blockchain/Web3 specialists. The alias pile is now official in one artifact: CL-STA-0240, DeceptiveDevelopment, DEV#POPPER, Famous Chollima, Gwisin Gang, PurpleBravo, Tenacious Pungsan, UNC5342, Void Dokkaebi, WaterPlum, the on-wiki Famous Chollima / REF9403 / BeaverTail-OtterCookie-WeaselBiscuit coverage all names this one operation.

Durable additions from the advisory read:

  • Formal organizational fusion: WaterPlum and some North Korean IT workers (PurpleDelta / Wagemole) are assessed to operate under the 313 General Bureau of the Munitions Industry Department, and the two clusters are "deeply intertwined", same IPs accessing laptop farms AND applying for jobs at Japanese crypto exchanges. This corroborates (from the state side) the credential-pipeline linkage Unit 42's token-jacking work (on-wiki) hypothesized between IT-worker operations and supply-chain malware.
  • Laptop-farm facilitation is now a named, dismantled artifact: a Japan-operated facilitator laptop farm was identified and dismantled; enablers in Japan, the U.S., and elsewhere run farms for remote device management; stolen victim ID images feed IT-worker impersonation (identity laundering is an output of the malware, not just a step).
  • Proxy-hiring moves to Discord (Silent Push, last week): a "Mouse Review" server recruiting U.S./EU/LATAM citizens as interview faces for $3,000-5,000, with the operator explicitly offering to remote-view the proxy's screen and complete live coding challenges, the human layer that beats geo/KYC controls. AI-generated ad text on record ("You handle communications and interviews. I handle all technical work behind the scenes.").
  • Kudelski Security (July 2026, on the internal infra): primary targets U.S. + Japan; Astrill VPN and Mullvad exit nodes for geo-look.
  • Defender translation: the campaign's success metric is a developer running the take-home project, every heuristic on this page still holds; new hunt add: interview-lane comms from Discord-recruited proxies means the first touch can be a legitimate-looking colleague/hiring-manager in your own pipeline. Screen external "code test" execution as untrusted code with zero credential reach.

Sources