Replication of Spracklen et al. (USENIX Security '25) on the 2026 frontier-LLM cohort.
"The Range Shrinks, the Threat Remains: Re-evaluating LLM Package Hallucinations on the 2026 Frontier-Model Cohort"
The paper reports a 127-name universal-hallucination set (109 PyPI, 18 npm), of which 53 (41 PyPI, 12 npm) remain registrable after each registry's existing defenses. Coordinated disclosure to PyPI Security and Socket.dev completed in April–May 2026.
The full enumerated list is no longer embargoed. An earlier revision of this
repository included it in a disclosure/ directory that was publicly
accessible from April 30, 2026; that directory has since been removed from
the current tree, but the list remains reachable in the tagged release
v0.2-preprint. The set is therefore treated as disclosed as of April 30,
2026.
Two caveats apply to the enumerated names, both discussed in §9.2 of the paper:
@ember/* npm names are virtual modules under
a controlled scope, ssh-keys is a published package, and metro-evaluator
is a security-hold placeholder.Aleksandr Churilov — churik0509@gmail.com
Replication of Spracklen et al. (USENIX Security '25) on the 2026 frontier-LLM cohort.
"The Range Shrinks, the Threat Remains: Re-evaluating LLM Package Hallucinations on the 2026 Frontier-Model Cohort"
The paper reports a 127-name universal-hallucination set (109 PyPI, 18 npm), of which 53 (41 PyPI, 12 npm) remain registrable after each registry's existing defenses. Coordinated disclosure to PyPI Security and Socket.dev completed in April–May 2026.
The full enumerated list is no longer embargoed. An earlier revision of this
repository included it in a disclosure/ directory that was publicly
accessible from April 30, 2026; that directory has since been removed from
the current tree, but the list remains reachable in the tagged release
v0.2-preprint. The set is therefore treated as disclosed as of April 30,
2026.
Two caveats apply to the enumerated names, both discussed in §9.2 of the paper:
@ember/* npm names are virtual modules under
a controlled scope, ssh-keys is a published package, and metro-evaluator
is a security-hold placeholder.Aleksandr Churilov — churik0509@gmail.com