Insights / AI and the software supply chain
AI and the software supply chain

When everyone vibe codes, one flaw hits everyone

In August, one hijacked GitHub account poisoned 444 npm packages in under four hours and went after the keys on every machine that installed one. AI coding tools make that kind of sameness normal. The tools are worth keeping. Code that ships unchecked is the risk.

Ausper Technology · Nov 18, 2026 · Insight · AI and the software supply chain · 9 min read

The short version
  • AI coding tools give thousands of teams the same packages and the same mistakes. Veracode found AI-generated code passes security tests 56% of the time, virtually unchanged.
  • One flaw now reaches everyone who shares it. ChainDrop poisoned 444 npm packages, about 2 billion monthly downloads, from one hijacked account in under four hours.
  • A patch does not revoke a stolen key. Find where AI-written code ships, at home and at your vendors, gate it with review, and rehearse revoking every key at once.

On August 4, attackers took over the GitHub account of one npm maintainer and used it to publish poisoned packages. StepSecurity, which named the worm ChainDrop, counted 444 packages and 2,212 versions poisoned in under four hours. Early reports ran higher, to 868 packages and past 1,300 (BleepingComputer). Together the affected packages see about 2 billion downloads a month (Aikido). On every machine that installed one, the malware went looking for GitHub and npm tokens, CI secrets, and credentials for AWS, Kubernetes, Vault, databases, Stripe, Slack, Azure and Google Cloud. StepSecurity’s advice to anyone who installed an affected version: “assume your system is compromised.” Nobody was targeted. They shared a dependency.

Fig. 1
ChainDrop kept spreading for three hours after the first alarm

August 4, 2026, in UTC. One hijacked maintainer account, one working day.

The attack
Poisoned commit pushed to the keyv repository09:02
First poisoned version published, with valid provenance09:35
Malicious releases surge: 444 packages09:40 to 13:20
The response
First public alarm10:17
npm starts unpublishing10:39
Main infected packages reverted18:10
Maintainer regains the accountabout 20:00
Sources: StepSecurity's ChainDrop timeline (UTC); CSO Online for the account recovery. Package count: StepSecurity and Aikido.
Data
ItemWhenNote
Poisoned commit pushed to the keyv repository09:02
First poisoned version published, with valid provenance09:35
Malicious releases surge: 444 packages09:40 to 13:20
First public alarm10:17
npm starts unpublishing10:39
Main infected packages reverted18:10
Maintainer regains the accountabout 20:00

AI coding tools make that kind of sharing the default. Your developers use them, and so do your vendors. The tools suggest the same packages and repeat the same mistakes for thousands of companies at once, so when one flaw is disclosed, everyone who shares it is exposed in the same hours. We use these tools ourselves and would not give them up. The exposure comes from code that ships without a check.

Speed is the feature. Sameness is the risk.

The tools got faster. The code they write did not get much safer. Veracode’s 2026 test of AI-generated code found an average security pass rate of 56 percent, which it calls “virtually unchanged” from last year. Java passed 30 percent of the time (Veracode, via SD Times).

The flaws are reaching real software. Georgia Tech’s Vibe Security Radar has confirmed 74 vulnerabilities (CVEs) traced to AI-generated code since it began counting in May 2025. The monthly count rose from 6 in January 2026 to 35 in March (Cloud Security Alliance). March alone had more than all of 2025 combined, Georgia Tech says.

Fig. 2
AI-linked CVEs went from 6 in January to 35 in March

Vulnerabilities traced to AI-generated code and confirmed by Georgia Tech's Vibe Security Radar, by month.

January 2026
6
February 2026
15
March 2026more than all of 2025 combined
35
Sources: Cloud Security Alliance (Apr 6, 2026); Infosecurity Magazine (Mar 26, 2026); Georgia Tech (Apr 9, 2026). 74 confirmed in all since May 2025.
Data
MonthCVEsNote
January 20266
February 202615
March 202635more than all of 2025 combined

The tools also invent dependencies. A 2025 study of 576,000 Python and JavaScript samples from 16 models (Spracklen and colleagues) found that 19.7 percent of the packages the models suggested did not exist (Cloud Security Alliance). An attacker who registers one of those names can be installed by every team whose assistant recommends it.

The platforms built for vibe coding have had the same trouble. At Lovable, an API flaw let anyone with a free account reach another user’s profile, public projects, source code and database credentials in as few as five API calls. It was still open 48 days after a researcher reported it (The Next Web). Moltbook, a vibe-coded social network, exposed 1.5 million API tokens within three days of launch through a database with no row-level security. Wiz found it and reported it, and Moltbook fixed it within hours.

The keys outlive the fix

When shared code fails, what leaks is usually a key. GitGuardian counted about 28.65 million new secrets pushed to public GitHub in 2025, up 34 percent. Secrets for AI services rose 81 percent. Commits made with one widely used AI coding assistant leaked secrets at about twice the baseline rate, 3.2 percent against 1.5. And more than 64 percent of the secrets that were valid in 2022 still worked in January 2026.

Fig. 3
Most of the secrets valid in 2022 still worked in 2026
64%+Secrets valid in 2022 that still worked when retested in January 2026.GitGuardian 2026
28.65MNew secrets pushed to public GitHub in 2025, up 34 percent.GitGuardian 2026
+81%Growth in leaked secrets for AI services, 2025 over 2024.GitGuardian 2026
3.2% vs 1.5%Leak rate of commits made with one AI coding assistant, against all public commits.GitGuardian 2026
Source: GitGuardian, The State of Secrets Sprawl 2026 (Mar 17, 2026).
Data
NumberWhat it isSource
64%+Secrets valid in 2022 that still worked when retested in January 2026.GitGuardian 2026
28.65MNew secrets pushed to public GitHub in 2025, up 34 percent.GitGuardian 2026
+81%Growth in leaked secrets for AI services, 2025 over 2024.GitGuardian 2026
3.2% vs 1.5%Leak rate of commits made with one AI coding assistant, against all public commits.GitGuardian 2026

That last number should worry a CEO most. A patch closes the hole. It does not take back a key that already left. On November 16 we counted 11 exploited flaws in AI tools added to CISA’s Known Exploited Vulnerabilities catalog this year, against 1 in 2025, and the lesson was the same: patching leaves leaked keys working. ChainDrop ran that lesson across a package registry. Every key it collected works until someone revokes it.

You may not vibe code. Your vendors do.

Most of the software a company runs, it did not write. Payroll, the CRM, the plugins in your build pipeline: another team built them. Your policy on AI coding covers your own developers. It does not cover the people who build what you buy.

Getting that wrong is expensive. IBM’s 2026 Cost of a Data Breach report ranks supply chain compromise as the second most common way breaches start, after phishing, and puts the average US breach at a record $11.5 million, nearly double the global average (eSecurity Planet has a summary).

Three questions to ask every vendor this quarter:

  • Which parts of what you ship us are written with AI tools, and what review does every change pass before release? You want a human reviewer and automated security tests on every change, whoever or whatever wrote it.
  • Can you list every package in each release, and tell us within a day whether you shipped one named in an advisory? A software bill of materials (SBOM) for each release is how a vendor answers that quickly.
  • Which of our credentials do you hold, and how fast can you revoke and replace all of them? Ask for the inventory, the owner of each key and the last time they rehearsed it.

The clock is hours now

In July, Microsoft fixed at least 570 flaws in a single Patch Tuesday, almost triple June’s count (Krebs on Security). Only 9 percent of organizations patch high-severity flaws within 24 hours, according to a Cloud Security Alliance survey of more than 900 practitioners. In June, an Ivanti Sentry flaw, CVE-2026-10520, saw confirmed exploitation within 40 hours of disclosure, the CSA reports. It was the first flaw to fall under CISA’s Binding Operational Directive 26-04 (BleepingComputer), which gives federal civilian agencies as little as three days to fix the riskiest known-exploited flaws (SecurityWeek).

Put those together. A shared flaw lands everywhere at once, an attacker can be working within two days, and most organizations take longer than a day to patch even when they know where to look. If you first have to find out which systems carry the flaw, you start behind. ChainDrop showed the same clock from the other side: the first public alarm came 75 minutes after the poisoned commit, and malicious releases kept coming for three hours after that. In October we showed the KEV window falling from 21 days to 3; this is that clock, applied to code nobody on your team wrote.

Our forecast: in 2027, at least one npm compromise will poison more packages than ChainDrop’s 444, by the count of the firm that discovers it. Shared dependencies give an attacker the most installs for one stolen account. If it misses, we will say so here.

Six things to check this quarter

  • Find where AI-written code ships. Ask each team and each vendor which assistants they use and where the output lands. You cannot gate what you cannot see.
  • Gate it like any other code. Code review and automated security tests on every change. At a 56 percent pass rate, the test catches what the tool misses.
  • Lock what you install. Pin versions, pull packages through a registry you control, and hold back names that are brand new or that nobody has used before. One in five suggested packages in that study did not exist.
  • Know who depends on what, within the hour. Keep an SBOM for every system, so when the next ChainDrop has a name you can list every system that pulled it the same morning.
  • Shorten and inventory your keys. Scan code and pipelines for secrets, swap long-lived tokens for short-lived ones, and rehearse revoking all of them at once.
  • Time your own patch clock. Measure the hours from disclosure to fix for high-severity flaws, and set the number beside 40 hours and BOD 26-04’s three days.

None of this asks you to give up the speed. Each item is a control your teams already run on code people write. The work is applying it every time, to your vendors as well as your own developers, and having your list ready before the next flaw has a name.

Ausper is positioned to help with the parts that take practice: security engineering, vulnerability management and CI/CD pipeline strategy, penetration testing and independent security assessment, and a security operations center that runs 24/7. We also build governed AI with human review in the loop, the same discipline AI-written code needs.

When the next shared flaw drops, will you be the cost, or the one still running? If you want a second set of eyes on the six, our free 30-minute AI exposure review walks through them with you.

Common questions

What is vibe coding?
Building software by describing what you want to an AI assistant and accepting much of what it writes with little line-by-line review. The same assistants also sit inside ordinary engineering teams, which is why the risk reaches companies that never use the term.
Should we ban AI coding tools?
We would not. The tools are fast, and your vendors will keep using them whatever your policy says. Put the same gates on AI-written code that you put on any code: review, automated security tests, pinned dependencies and short-lived keys. Veracode’s 2026 tests found AI-generated code passed security tests 56 percent of the time, so the gate does the work.
What was ChainDrop?
An npm supply-chain worm found on August 4, 2026. Attackers took over a maintainer’s GitHub account and pushed poisoned releases that spread to 444 packages and 2,212 versions in under four hours, by StepSecurity’s count, with about 2 billion monthly downloads between them. Early press reports gave higher counts. The malware collected developer and cloud credentials from machines that installed an affected version.
What should we ask our vendors?
Three things. Which parts of what they ship you are written with AI tools, and what review every change passes before release. Whether they can list every package in each release and tell you within a day if they shipped one named in an advisory. And which of your credentials they hold, and how fast they can revoke and replace all of them.
What is the AI exposure review?
A free 30-minute conversation with our team about where AI-written code, shared packages and long-lived keys could expose you and your vendors, and what we would fix first. We never ask for code, credentials or system access on the call. Book it here.

Sources

Last verified Nov 18, 2026

  1. StepSecurity, ChainDrop npm Worm: Bun-loaded CI/CD credential harvester with Ethereum dead-drop C2 (Aug 4, 2026)Named the worm. 444 packages and 2,212 versions poisoned in under four hours; the hour-by-hour timeline in UTC; "assume your system is compromised." Accessed Nov 18, 2026.
  2. Aikido Security, Keyv and friends compromised in active supply chain attack (Aug 4 to 5, 2026)The maintainer's GitHub account was compromised on August 4; at least 444 packages across 1,381 versions; over 2 billion monthly installs; the credentials the malware collects. Accessed Nov 18, 2026.
  3. BleepingComputer, Massive ChainDrop npm supply-chain attack infects hundreds of packages (Aug 4, 2026)Early counts of 868 and more than 1,300 packages; the credentials collected, including Azure and GCP. Accessed Nov 18, 2026.
  4. CSO Online, ChainDrop credential-stealing worm infects over 400 npm packages (Aug 4, 2026)The developer regained the GitHub account at around 20:00 UTC. Accessed Nov 18, 2026.
  5. Veracode, 2026 GenAI Code Security Report, press release via SD Times newswire (Jul 28, 2026)Average security pass rate 56 percent, "virtually unchanged"; Java 30 percent; Python 63 percent; no security-specific prompting. Accessed Nov 18, 2026.
  6. Veracode, 2026 GenAI Code Security Report: AI risk (blog, Jul 28, 2026)56 percent, "barely changed from 55% in the first report." Accessed Nov 18, 2026.
  7. Cloud Security Alliance, Vibe Coding Security Debt: AI-Generated Vulnerabilities at Scale (research note, Apr 6, 2026)Georgia Tech's Vibe Security Radar: 74 AI-linked CVEs through March 2026, 6 in January to 35 in March; the 576,000-sample study: 19.7 percent of suggested package dependencies hallucinated. Accessed Nov 18, 2026.
  8. Georgia Tech College of Computing, Bad vibes: AI-generated code is vulnerable, researchers warn (Apr 9, 2026)About 18 cases in the second half of 2025; 56 in the first three months of 2026; March alone 35, "more than all of 2025 combined." Accessed Nov 18, 2026.
  9. Infosecurity Magazine, AI-generated code vulnerabilities (Mar 26, 2026)35 in March, up from six in January and 15 in February. Accessed Nov 18, 2026.
  10. Spracklen et al., We Have a Package for You! A Comprehensive Analysis of Package Hallucinations by Code Generating LLMs (USENIX Security 2025; arXiv:2406.10279)The study behind the hallucinated-package figure: 576,000 code samples from 16 models. Accessed Nov 18, 2026.
  11. The Next Web, Lovable left thousands of projects exposed for 48 days (Apr 21, 2026)Lovable: any free account could reach another user's profile, public projects, source code and database credentials in as few as five API calls; open 48 days. Moltbook: breached within three days of launch, 1.5 million API tokens, no row-level security. Accessed Nov 18, 2026.
  12. Wiz, Exposed Moltbook database reveals millions of API keys (Feb 2, 2026)Wiz found and disclosed the exposure; Moltbook "secured it within hours." Accessed Nov 18, 2026.
  13. GitGuardian, The State of Secrets Sprawl 2026 (Mar 17, 2026)28.65 million new secrets on public GitHub in 2025, up 34 percent; AI-service secrets up 81 percent; 3.2 percent leak rate for commits made with one AI coding assistant against a 1.5 percent baseline; above 64 percent of 2022's valid secrets still valid in January 2026. Accessed Nov 18, 2026.
  14. IBM, Cost of a Data Breach Report 2026US average breach cost a record USD 11.5 million, nearly double the global average (p. 9); supply chain compromise the second most common initial attack vector (p. 21). Accessed Nov 18, 2026.
  15. eSecurity Planet, IBM 2026 Cost of a Data Breach Report: Key Findings (Jul 30, 2026)Supply chain compromises ranked second, after phishing; US average $11.5 million. Accessed Nov 18, 2026.
  16. Cloud Security Alliance, AI Vulnerability Discovery Is Outpacing Patch Capacity (research note, Aug 2, 2026)570 fixes in July, nearly triple June; 9 percent of organizations patch high-severity flaws within 24 hours (CSA survey of more than 900 practitioners); Ivanti Sentry CVE-2026-10520 exploited within 40 hours of disclosure. The note says it was drafted with AI assistance and has not had formal CSA review, so we also cite the reports behind it. Accessed Nov 18, 2026.
  17. Krebs on Security, Microsoft patches a record 570 security flaws (Jul 14, 2026)At least 570 fixes, almost triple the month before. Accessed Nov 18, 2026.
  18. BleepingComputer, CISA gives feds 3 days to patch Ivanti flaw exploited in attacks (Jun 12, 2026)CVE-2026-10520 is the first vulnerability BOD 26-04 applies to. Accessed Nov 18, 2026.
  19. SecurityWeek, CISA directs federal agencies to prioritize security patches based on risk (Jun 11, 2026)BOD 26-04: three days for the riskiest known-exploited flaws on exposed assets, then 14 or 60 days. CISA's own page refused our automated request. Accessed Nov 18, 2026.

Get Insights by email

Thanks for subscribing.

Share on LinkedInFollow AusperSubscribe by RSS

Related reading

Free 30-minute check

Where would one shared flaw hit you?

Thirty minutes with our team on your AI-written code, your vendors, your keys and your patch clock. No code, credentials or access needed.

Book the AI exposure review