LLMs Gone Wild - AI-Crafted Malware Infecting Our Supply Chain

The talk opens with how much of a modern application is built from open-source packages the team never chose directly. Transitive dependencies concentrate risk in widely used projects, and a single overworked maintainer can become a single point of failure. The speaker uses the Debug and Chalk package compromise, where a phished maintainer account pushed malicious code into libraries with billions of weekly downloads, to show how the usual CVE-based process falls behind malware that spreads and disappears within hours.

The middle section covers three ways attackers use AI: as a tool running on the victim's own machine, as a productivity boost for producing convincing lures and bloated packages, and as a generator of polymorphic payloads that defeat signature matching. Examples include the Nx s1ngularity malware, which relied on locally installed AI clients to hunt for credentials, and hallucinated package names that attackers can register once language models keep recommending them. Skill registries for an embedded AI agent are presented as another route for malicious code to reach users.

The closing part describes a detection approach that runs rule-based scans on every incoming package, passes the flagged indicators to a model, and then relies on a human to make the final call. Case studies such as the Ripple XRP SDK compromise and invisible Unicode payloads illustrate the range of concealment tricks seen in practice. The main takeaway is that deterministic rules alone no longer keep pace, and teams should expect malicious releases to arrive faster than manual advisories can track.