The Zero Day Initiative, the bug-bounty programme run by TrendAI, published 47 advisories on 11 November covering Meta’s open-source AI research libraries, every one labelled “0Day” and rated 7.8 on the CVSS severity scale. They cover fairseq (31), FAIRChem (9), Detectron2 (6) and fairseq2 (1).
ZDI’s vendor timelines show 43 of the reports reached Meta on 26 March 2026 and 4 on 4 March. By our arithmetic, that is 230 days from report to publication for 43 of the 47, and 252 days for the other 4. Seventeen FAIRChem flaws ZDI reported the same March day were fixed within 5 days.
Business Pill · Open Weights Explained
An evergreen Business Engineer explainer on renting a model through an API versus running open code and weights yourself, which puts maintenance on the user. It teaches the general idea only and does not discuss the advisories.
The key insight: As we read the record, the same Meta library shows both outcomes: 17 FAIRChem flaws reported on 26 March were fixed within 5 days, while 9 FAIRChem flaws from the same March reports were published as zero-days on 11 November.
International Cyber Digest flagged the batch on X at 03:04 UTC on 12 November: “Meta left 47 code-execution bugs in its open-source AI libraries unpatched since March, so bug-bounty programme Zero Day Initiative has published them all as zero-days.” We checked the claim against ZDI’s published-advisory list and all 47 advisory pages.
Each advisory describes a remote code execution flaw: deserialization of untrusted data or code injection, mostly in how the libraries load model checkpoints and configuration files. In ZDI’s words, “User interaction is required to exploit this vulnerability in that the target must visit a malicious page or open a malicious file.” We do not describe the technical detail here.
Zero Day Initiative advisories for Meta’s open-source AI libraries by library and status: 17 FAIRChem flaws fixed (advisories ZDI-26-793 to 809, published 26 October 2026, pointing to fairchem pull request 1935, merged 31 March) and 47 published as 0-day on 11 November 2026 (ZDI-26-907 to 953), with Meta’s replies as recorded in ZDI’s vendor timelines. Counts are our count of the ZDI pages.
What Meta Told ZDI
ZDI records each vendor reply in a dated timeline. For 32 of the 47 advisories (31 fairseq and the one fairseq2), Meta replied that the fairseq repository is no longer maintained and was archived; 30 of those replies are dated 17 September 2026 and 2 are dated 7 April. On the two April timelines, ZDI added: “However, GitHub shows the repository was actually archived on March 20, 2026, after this report was submitted”.
One Detectron2 report was answered as “a duplicate of another issue”, on 6 March. One FAIRChem report drew the reply, on 13 October, that the reported issue “does not have a security/privacy impact”. For the other 13, the timelines record no vendor communication beyond acknowledging receipt, or in one case that the report was being investigated. Most timelines show ZDI asking for updates on 27 July, and all 47 record its notice of the intention to publish on 23 October.
Each of the 47 advisories gives the same mitigation: “Given the nature of the vulnerability, the only salient mitigation strategy is to restrict interaction with the product”.
The 17 That Were Fixed
The contrast sits inside FAIRChem. On 26 October, ZDI published 17 FAIRChem advisories, ZDI-26-793 to ZDI-26-809, also rated 7.8 and also reported on 26 March. Each states: “Meta has issued an update to correct this vulnerability.” They point to pull request 1935, titled “mitigate unsafe pickle deserialization”, which GitHub shows was merged on 31 March 2026.
By our count, that makes 64 ZDI advisories for Meta since 26 October: 17 fixed and 47 published without a fix. Nine of the 47 are in FAIRChem itself, the same library.
Our Numbers
By our arithmetic, 32 of the 47 (68%) drew the reply that the repository was archived, and 13 (28%) drew no substantive reply. The 17 fixed FAIRChem flaws took 5 days from report to merged fix; the 47 zero-days were published 230 to 252 days after the reports.
The libraries are widely used. The GitHub API on 12 November lists fairseq as archived with 32,207 stars and 6,664 forks; Detectron2, not archived, with 34,984 stars and 7,938 forks; FAIRChem with 2,356 stars; and fairseq2, not archived and last pushed to on 21 October, with 1,154. By our arithmetic, the four repositories carry 70,701 stars and 15,279 forks between them.
As we read the timelines, the archive reply covers 32 of the 47 advisories, which moves the patching question from Meta to whoever still runs fairseq: ZDI’s stated mitigation is to restrict interaction with the product.
As we read the GitHub figures, the exposure is not a niche one: by our arithmetic, the four repositories carry 70,701 stars and 15,279 forks, and three of the four are not archived.
Zero Day Initiative, advisory ZDI-26-953, 11 November 2026
“the vendor communicated that the reported issue does not have a security/privacy impact”
Three Implications
FOR COMPANIES RUNNING FAIRSEQMeta replied to ZDI that the fairseq repository is no longer maintained and was archived; ZDI’s stated mitigation is to restrict interaction with the product.
FOR DETECTRON2 AND FAIRCHEM USERS ZDI published 15 advisories for the two libraries on 11 November, 6 for Detectron2 and 9 for FAIRChem, each rated 7.8.
FOR SECURITY TEAMS Each advisory says user interaction is required: the target must visit a malicious page or open a malicious file.
The framework sets out the pattern: “Close the SCARCE layer (keep proprietary) + Open the ABUNDANT layer (commoditize)”.
As we read it, an opened research library is the abundant layer, and this record shows its upkeep is a separate decision: one library fixed in 5 days, another archived. The framework predates the advisories and does not discuss them.
What Is Not Established
The advisories do not say whether any of the 47 flaws have been exploited. ZDI’s pages are the record of Meta’s replies; we read no statement from Meta beyond what those timelines quote, and we did not contact Meta for this piece.
The advisories do not state a fix date for the Detectron2, FAIRChem or fairseq2 flaws. One FAIRChem advisory lists a report date of 24 February 2026 in its disclosure timeline and 4 March in its vendor timeline; we use the vendor timeline.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
ZDI published 47 zero-day advisories for Meta’s open-source AI libraries on 11 November, 230 to 252 days after the reports, by our arithmetic. For 32 of them Meta pointed to an archived repository; 17 sister flaws in FAIRChem were fixed within 5 days in March. For companies that run these libraries, ZDI’s stated mitigation is to restrict interaction with them.
A note on sourcing. We read the Zero Day Initiative’s published-advisory list and all 47 Meta advisory pages of 11 November 2026, plus the 17 FAIRChem advisories of 26 October; Meta’s replies quoted here are as recorded in ZDI’s vendor timelines. We read fairchem pull request 1935 on GitHub and took repository figures from the GitHub API on 12 November 2026. The lead came from an International Cyber Digest post on X.
Day counts, shares and totals are our arithmetic. We did not contact Meta or ZDI for this piece, and we do not publish technical exploitation detail.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
