News · DiggingBeagle record

Fake Web3 accelerator shipped a coding agent with hidden routing and made-up audits

Elastic traced an AI-enabled developer-tool lure to malicious packages, external model-routing changes and 18 fictitious MCP audits.

Published
Oct 14, 2026

The report

Elastic Security Labs traced the purported Apex Accelerator, which promised investment connections and AI credits, to npm packages that retrieved additional executables. The operators packaged a coding assistant with modified model-routing settings and a remote MCP server. The study also inspected a supposed contract-auditing package whose 18 security tools emitted fabricated results rather than carrying out scans.

The scheme joined three different forms of authority confusion. A package installation could execute platform-specific code, the coding agent could submit sensitive material to an undisclosed endpoint, and an apparently automated security report could be accepted despite no actual inspection. Elastic demonstrated these mechanisms in controlled analysis but did not establish how many real customers installed the packages or lost funds.

The technical dossier separates those observed mechanisms from unverified victim outcomes. This is a genuine software supply-chain attack against AI developers, not evidence that every remote MCP tool or security assistant is malicious.

The fake audit is not merely a low-quality recommendation. According to Elastic, its advertised checks were replaced by fabricated output, while the package-distribution route created a separate opportunity to execute code. Neither observation determines how many developers actually installed or trusted the product.

What could be verified

The code and tool-handler examination supports three separate observations: installation-time downloads, routing controlled by the distributed assistant, and audit functions returning values without executing their advertised tests. Only the last establishes that a purported audit was fabricated. A generated score cannot confirm that a contract is safe, and no amount of fluent explanation substitutes for the command or trace of an actual scan. The open questions are the number of installs that completed their platform-specific bootstrap, the data actually transmitted by those installs, and whether any recipient acted on the fictitious findings.

The distinction also determines remediation. Removing the npm package does not automatically revoke credentials sent to an unexpected endpoint, nor does it correct past project decisions based on synthetic audit reports. The case evidence should be consulted for each observed component before attributing an outcome to the campaign.

Research behind this