Quick answer
AI coding assistants speed up development but introduce new risks: insecure code suggestions accepted without review, secrets sent to or reproduced by the assistant, prompt injection through repository content, agentic assistants with access to terminals and credentials, and slopsquatting, where attackers register package names that assistants hallucinate.
On this page
How It Works
Assistants generate code from patterns, including insecure ones. Agentic coding tools can run commands, install packages and edit files, so instructions hidden in a README, issue or dependency can steer them. When an assistant suggests a package that does not exist, attackers who register that name can have malicious code installed automatically. This is known as slopsquatting.
Why It Matters
Code ships faster, with less human scrutiny, and assistants often run with the developer's own credentials and access.
How to Test for It
- Review which assistants are in use and what access agentic tools have to terminals, secrets and repositories.
- Test whether repository content can inject instructions into agentic assistants.
- Scan for secrets in code and assistant configuration.
- Check that dependency installation is controlled and verified.
- Include AI-generated code in normal code review and security testing.
How to Reduce the Risk
Approve assistants centrally, restrict agentic tools to sandboxes without production credentials, require human review of changes, use dependency allow-lists and verification, scan for secrets and keep application security testing in the release process.
Regulatory and Framework Context
For Australian organisations, governing AI coding tools is part of managing shadow AI and supply chain risk under ASD's secure-by-design guidance.
Being exploited now
Threat Radar →The latest related vulnerabilities confirmed as exploited in the wild.
- CVE-2026-97229·3 weeks ago·Dev toolsGitLab Community Edition and Enterprise EditionGitLab Community Edition and Enterprise Edition Integer Overflow or Wraparound VulnerabilityEPSS 69.3%
- CVE-2026-21589·5 weeks ago·Dev toolsAtlassian Multiple ProductsAtlassian Multiple Products Files or Directories Accessible to External Parties VulnerabilityEPSS 99.7%
- CVE-2026-102489·7 weeks ago·Dev toolsZammad GmbH ZammadZammad GmbH Zammad Session Fixation VulnerabilityEPSS 49.9%
- CVE-2026-102490·7 weeks ago·Dev toolsZammad GmbH ZammadZammad GmbH Zammad Improper Privilege Management VulnerabilityEPSS 10.9%
Latest intel on this topic
- Borrowed Code, Real Consequences: What OPTIMJACK Teaches UsUnit 42 · 17 November 2026
- The risks of device code authentication for AI coding agentsMicrosoft Security Blog · 17 November 2026
- The US wants its quantum lead to turn into factories and supply chainsCyberScoop · 10 November 2026
- Why Europe's Freezer Hacks Spread So Far, So FastDark Reading · 6 November 2026
- Updated guidance on the 2026 Minimum Elements for a Software Bill of MaterialsASD's ACSC · 2 September 2026
Updated hourly from the daily briefing.
Common questions
What is slopsquatting?
Registering software package names that AI assistants tend to hallucinate, so developers who install the suggestion get the attacker's code.
Is code from AI assistants less secure?
It can contain the same flaws as human code, and it is often reviewed less carefully. Treat it like any other untrusted contribution.
Change log
- 13 November 2026First published.