A Vulnerable Package and a Malicious Package Are Different Problems

If a build of yours installed axios 1.14.1 on March 31, 2026, matching CVEs against your dependencies would not have caught it. The OSV record for that release lists one alias, a GitHub advisory, and no CVE. The install itself ran a postinstall hook leading to multi-stage payloads including a remote access trojan.
Most dependency programs ask one question: which CVEs affect what we run? That is the right question for a flaw and the wrong one for a poisoned release. The two need different signals, different clocks and different responses, and a single triage queue handles the second one badly.
Why "a very bad vulnerability" fails
The obvious move is to treat a poisoned release as a very bad vulnerability: wait for the advisory, match it to your inventory, rank it, patch it. Each step assumes something that is not true here.
NIST defines a vulnerability as a weakness "in an information system, system security procedures, internal controls, or implementation that could be exploited or triggered by a threat source". No flaw in axios's code was exploited. The Register reported that the attackers left axios's own code alone and added a separate package, plain-crypto-js, to deliver the payload. The project's post-mortem says they got in through the lead maintainer's PC, using targeted social engineering and RAT malware. It also names process gaps, such as publishing directly from a personal account. Those are real, but they sit in how releases get published, not in code you could patch. LiteLLM says a maintainer's PyPI account may have been compromised, and that stolen credentials were reportedly used to reach its publishing pipeline. The release did what its publisher wanted. The publisher was not who you trusted.
Ranking breaks for the same reason. CVSS, EPSS and KEV help you decide which flaws deserve attention first. The GitHub advisory for the malicious LiteLLM releases is marked CRITICAL, which is fair and changes nothing. The question is whether that code ran on your machine.
The clock runs somewhere else
The post-mortem's timeline for axios: 1.14.1 published at 00:21 UTC, 0.30.4 and the first external detections around 01:00, both versions removed at 03:15. Google's threat intelligence group puts the end at 03:20. The OSV record's published timestamp is 03:15:49, the same minute the versions came down. Anyone querying that record during the window got nothing.
LiteLLM has the same shape with messier edges. The project says 1.82.7 and 1.82.8 were live on March 24 from 10:39 UTC "for about 40 minutes" before PyPI quarantined them, while PyPI's April 2 incident report puts upload to quarantine at 2h 32m. The project's page also asks you to assess impact for 10:39 to 16:00 UTC, so use that wider window. The PyPI advisory record was published at 15:35 UTC, and GitHub's the next day. That record says the code runs on importing any module, while the incident report says it ran on install. axios ran at install. In both cases the code does not wait for an attacker to reach anything.
Now the contrast, in the same package. On April 3, GHSA-53mr-6c8q-9789, also CVE-2026-35029, described a privilege escalation in LiteLLM, fixed in v1.83.0. On April 20, the project's GHSA-r75f-5x8p-qvmc described an unauthenticated SQL injection in proxy API key verification, fixed in 1.83.7. One project, one month, two kinds of record. OpenSSF drew the same line: its malicious packages repository puts vulnerability reports out of scope, and the OSV schema gives it a separate MAL prefix.
| Vulnerable package | Malicious package | |
|---|---|---|
| What it is | A flaw nobody intended | Code somebody intended, under the real name |
| Signal | Advisory with CVE, affected range, fix version | Malware record; the ones for these axios and LiteLLM releases list no CVE |
| Clock | From disclosure until you patch | From publication until removal |
| The question | Is it likely to be exploited here? | Did anything install or import it in the window? |
| What ends it | Upgrading | Not removal: you assume compromise and rotate |
The response is incident response
That last row matters most. Upgrading closes a vulnerability. The axios record says any computer with the package "should be considered fully compromised", that secrets should be rotated from a different computer, and that removal comes with "no guarantee". CISA's alert says to go backward to axios@1.14.0 or axios@0.30.3, delete node_modules/plain-crypto-js/, and rotate or revoke exposed credentials. That alert came on April 20, twenty days after removal, and it was still useful, because this work looks backward.
The Shai-Hulud worm made the same point in September 2025. CISA reported over 500 npm packages compromised and advised pinning to releases from before September 16, rotating developer credentials and phishing-resistant MFA.
So route these to whoever handles incidents, not to the patch backlog. A patch SLA means nothing for a cloud key that may already be gone.
Prevention pulls the other way
For npm, CISA's alert recommends these two .npmrc settings:
ignore-scripts=true
min-release-age=7
A seven-day wait outlasts a three-hour window, and the first setting removes the postinstall route. Neither touches import-time code on PyPI. And the cooldown collides with vulnerability work, which rewards reaching a fixed release quickly. Keep the wait as your default and decide now how a security fix gets past it, with a named reviewer.
The post-mortem also says that if you were already pinned to a clean version and did not run a fresh install in the window, you are fine.
Where this stops applying
None of this is an argument for dropping CVE work. Verizon's 2025 DBIR has vulnerability exploitation as the initial access step in 20% of breaches and third parties involved in 30%, a figure not specific to malicious packages.
Nor do databases ignore malware. GitHub and OSV both list it, and those records are how you check afterwards. They are detection after publication, not prevention.
Download totals also need care. Google's figure of over 100 million weekly axios downloads is volume, not installs in the window. For LiteLLM, PyPI counted over 119k downloads of the bad versions during the window. Expect exposure to concentrate in unpinned installs, CI that resolves ranges, and developer machines.
The two problems do share one thing: an inventory. A malicious release asks more of it, though. Knowing you run litellm today does not tell you what was installed at 11:00 UTC on March 24.
Run the March check on your own systems
Search every lockfile and requirements file you own for axios 1.14.1 or 0.30.4, plain-crypto-js, and litellm 1.82.7 or 1.82.8. Then do the harder half: search CI logs, image build logs and artifact caches for installs between 00:21 and 03:20 UTC on March 31 and between 10:39 and 16:00 UTC on March 24. Include developer laptops if you can reach them.
Write down what you could not answer. Purged logs or an uncovered team is your real finding, and the next poisoned release will land in the same gap.
References
- CISA, Supply Chain Compromise Impacts Axios Node Package Manager, April 20, 2026
- axios project, Post Mortem: axios npm supply chain compromise, April 2, 2026
- Google Threat Intelligence Group, North Korea-Nexus Threat Actor Compromises Widely Used Axios NPM Package, April 1, 2026
- The Register, Top npm package backdoored to drop dirty RAT on dev machines, March 31, 2026
- OSV.dev and OpenSSF, MAL-2026-2307, March 31, 2026
- LiteLLM, Security Update: Suspected Supply Chain Incident, March 24, 2026, updated through March 30, 2026
- PyPI, Incident Report: LiteLLM/Telnyx supply-chain attacks, with guidance, April 2, 2026
- GitHub Advisory Database via OSV.dev, GHSA-5mg7-485q-xm76, March 25, 2026
- PyPI Advisory Database via OSV.dev, PYSEC-2026-2, March 24, 2026
- LiteLLM, GHSA-53mr-6c8q-9789, April 3, 2026
- LiteLLM, GHSA-r75f-5x8p-qvmc, April 20, 2026, updated April 28, 2026
- CISA, Widespread Supply Chain Compromise Impacting npm Ecosystem, September 23, 2025
- OpenSSF, Malicious Packages repository, README last changed September 30, 2025
- OpenSSF, Open Source Vulnerability schema, schema document last changed March 31, 2026
- NIST CSRC, Glossary: vulnerability, as archived March 28, 2026
- Verizon, 2025 Data Breach Investigations Report, April 23, 2025
