Inside the Open-Source Kill Chain: How LLMs Helped Catch Lazarus and Stop a Crypto Backdoor

Mackenzie Jackson opens with the scale of open-source risk: most of an application's code comes from packages maintained by volunteers, so a compromise or hidden flaw can reach very large numbers of downstream users. Existing defenses depend on CVE disclosure, which is slow and often skipped, so maintainers can quietly fix serious bugs without anyone outside the project being told.

The middle of the talk describes a project that scans changelogs from millions of packages with language models. The models standardise inconsistent entries and flag security-related changes that keyword rules would miss. The pipeline reported hundreds of undisclosed vulnerabilities, most of which never received a CVE. A second pipeline pairs traditional scanners with an LLM that ranks their noisy output for human review. Examples include an account takeover of a widely used parsing library, a Lazarus payload hidden with whitespace in a typosquatted package, and a backdoor in an official Ripple SDK that could have exposed wallet keys at exchanges. A heavily obfuscated package hid its payload behind invisible Unicode characters and chained Base64 links.

The closing takeaway is that LLMs work best as triage and context layers alongside detectors, not as standalone scanners. The speaker noted that the data and a wrapper tool that blocks known malicious installs are free to use, and the Q&A covered model choice, maintainer responsiveness and falling costs.