An AI coding assistant can be turned against its own user without any model being hacked. In November 2026, Bitdefender Labs reported 22 malicious GitHub links whose project files told coding agents to run a hidden script that stole developer credentials. This checklist turns that case into controls: a core list for any organisation, then add-ons for healthcare in the EU and for banking, financial services and insurance (BFSI) in India. Each check names the standard or rule that supports it, and some go beyond the standards.
Last reviewed 19 November 2026.
What happened
Files that coding agents read for project guidance, such as AGENTS.md and CLAUDE.md, told the agent to run a setup script. The same files called the script signed and told the agent to hide its output and not inspect it. A second path used editor and agent settings files that run commands when a folder opens, with no prompt at all.
The payload collected GitHub and GitLab tokens, cloud keys, SSH keys, .env files, API keys and saved browser passwords. It then checked which stolen tokens could write to repositories, committed the same poisoned files into up to 25 more repositories, and edited agent settings on the machine for 14 named coding tools.
Why it worked
Three conditions met on one laptop:
- The agent treated text in the project as instructions, not data. This is indirect prompt injection.
- The agent could run commands without a person approving them.
- The agent ran with the developer’s own credentials, so it could reach everything the developer could.
Remove any one of the three and the attack fails or shrinks. The checklist is built around that idea. Our guide to the nine AI attacks in ISO/IEC 27090 explains prompt injection and the other eight threats in more depth.
Core checklist for every organisation
Tick each item when it is in place. The reference in italics shows where a standard supports the check.
1. Govern
- ☐ Set the risk appetite for AI agents that run code or reach secrets, at board level. IEEE 2863 5.2.3.4 f), ISO/IEC 42001 A.2.2
- ☐ Name an owner for each coding assistant in use. ISO/IEC 42001 A.3.2
- ☐ Write a rule for approved coding assistants and approved uses. ISO/IEC 42001 A.9.3, A.9.4
2. Inventory and assess
- ☐ Add every coding assistant, plugin and MCP server to the AI inventory. ISO/IEC 42001 4.3, A.4.2
- ☐ Assess each one in the risk register: what it can read, run and reach. ISO/IEC 42001 6.1.2, ISO/IEC 27001 8.2
- ☐ Assess each vendor: data retention, training use, hosting location and incident notice. ISO/IEC 42001 A.10.3, ISO/IEC 27001 A.5.19 to A.5.21, ISO/IEC 27090 5.5
3. Configure the agent
- ☐ Keep approval prompts on for shell commands, scripts and package installs. ISO/IEC 27090 6.11.3.2
- ☐ Run agents in a sandbox or container with no path to production. ISO/IEC 27090 6.11.3.3, ISO/IEC 27001 A.8.31
- ☐ Allow-list the commands and network destinations each agent may use. ISO/IEC 27090 6.11.3.3
- ☐ Turn off auto-run tasks and session hooks for folders not yet trusted. ISO/IEC 27001 A.8.9. Beyond the standards
- ☐ Approve MCP servers and plugins centrally, and block unknown ones. ISO/IEC 27001 A.8.19, ISO/IEC 42001 A.10.3
4. Credentials
- ☐ Give each agent its own identity, with short-lived tokens scoped to the task. ISO/IEC 27090 6.11.3.3, ISO/IEC 27001 A.5.17, A.8.2
- ☐ Keep secrets out of workspaces: no keys, tokens or .env files in repositories. ISO/IEC 27001 A.5.17, A.8.4
- ☐ Issue fine-grained repository tokens, without workflow rights unless the task needs them. Beyond the standards
5. Review what the agent reads and writes
- ☐ Treat agent instruction files as code: owner review and protected branches for AGENTS.md, CLAUDE.md, .cursorrules, .mcp.json and .vscode/tasks.json. ISO/IEC 27001 A.8.4, A.8.32
- ☐ Inspect unfamiliar repositories before you open them in an AI-enabled editor. Beyond the standards
- ☐ Flag any instruction that says hide output, do not inspect, or skip confirmation. ISO/IEC 27090 7.7
- ☐ Review and test AI-written code like any other code: peer review, static analysis and dependency checks. ISO/IEC 27001 A.8.25, A.8.28, A.8.29, ISO/IEC 42001 A.6.2.4
- ☐ Pin and verify every dependency an agent adds. ISO/IEC 27090 5.5.4
6. Monitor
- ☐ Log every agent command, tool call and file change, and keep the logs off the developer machine. ISO/IEC 27090 7.5, ISO/IEC 42001 A.6.2.8, ISO/IEC 27001 A.8.15
- ☐ Alert when an agent touches credential stores, SSH keys or browser data. ISO/IEC 27001 A.8.16
7. Respond
- ☐ Add a compromised AI agent scenario to the incident playbook. ISO/IEC 27001 A.5.24, A.5.26, ISO/IEC 42001 A.8.4
- ☐ If a poisoned file ran, isolate the machine, rotate every token and key it could reach, and check repositories for new commits. Beyond the standards
- ☐ Know your reporting clocks: CERT-In within 6 hours, and DPDP Rule 7 from 13 May 2027 when personal data is affected. CERT-In Directions 2022, DPDP Rules 2025
8. People and testing
- ☐ Train developers on indirect prompt injection, using this case. ISO/IEC 42001 7.2, 7.3, ISO/IEC 27001 A.6.3
- ☐ Red team the coding agent setup at least once a year, and report the results to the governance body. IEEE 2863 6.8, ISO/IEC 27090 5.7
Add-on: healthcare in the EU
Use these on top of the core checklist when you build health software or run a hospital in the EU. The coding assistant is not itself a high-risk AI system. The code it writes can still end up inside a regulated medical device.
If you build medical device or health software
- ☐ Run AI-written code through the same software life cycle as human code. IEC 62304
- ☐ Cover the coding assistant in your secure development process. Notified bodies expect IEC 81001-5-1 as evidence for MDR Annex I sections 17.2 and 17.4. EN IEC 81001-5-1, MDCG 2019-16
- ☐ Validate the coding assistant as software used in the quality management system, before use. ISO 13485 4.1.6
- ☐ Add the poisoned instruction file to the device risk file as a threat to the development chain. ISO 14971, MDR Annex I 17.2
- ☐ Record which code an AI tool wrote or changed, in the design and development records. MDR Annex II
- ☐ If the product is an AI medical device, plan for EU AI Act high-risk duties, including accuracy, robustness and cybersecurity. The Digital Omnibus provisional agreement of 7 May 2026 moves embedded high-risk duties to 2 August 2028. EU AI Act Article 15, EN 18286
If you are a hospital or other health entity
- ☐ Confirm whether you are an essential or important entity under NIS2, and get management approval for the measures. NIS2 Article 20, Article 21
- ☐ Apply supply chain security to coding assistant vendors, including their secure development practices. NIS2 Article 21(2)(d)
- ☐ Apply security in development and maintenance to every internal tool an agent helps build. NIS2 Article 21(2)(e)
- ☐ Prepare the reporting ladder for a significant incident: early warning within 24 hours, notification within 72 hours, final report within one month. NIS2 Article 23
For both
- ☐ Keep patient data out of prompts, repositories and agent workspaces. Use synthetic test data. GDPR Article 25, Article 32
- ☐ Check where the vendor processes prompts and code. A transfer outside the EU needs a GDPR transfer basis. GDPR Chapter V
- ☐ If personal data is breached, notify the supervisory authority within 72 hours. GDPR Article 33
Add-on: BFSI in India
Use these on top of the core checklist for banks, NBFCs, payment operators, market intermediaries and insurers. A SaaS coding assistant is usually an outsourced IT service, so the outsourcing rules apply to it.
Banks, NBFCs and other RBI regulated entities
- ☐ Bring coding assistants under the board-approved IT governance framework and its secure development rules. RBI Master Direction on IT Governance, Risk, Controls and Assurance Practices, 7 November 2023
- ☐ Treat a SaaS coding assistant as IT outsourcing: due diligence, contract terms and right to audit. RBI Master Direction on Outsourcing of IT Services, 10 April 2023
- ☐ Write into the contract that the vendor reports cyber incidents without delay, so you can report to RBI within 6 hours of the vendor detecting them. RBI Master Direction on Outsourcing of IT Services
- ☐ Keep payment system data out of prompts, repositories and agent workspaces. RBI requires it to be stored only in India. RBI circular on storage of payment system data, 6 April 2018
- ☐ List coding assistants in the AI inventory and the board-approved AI policy. Its recommendations are expected to flow into RBI directions. RBI FREE-AI framework, 13 August 2025
SEBI regulated entities
- ☐ Map coding assistant controls to your cyber resilience framework obligations. SEBI CSCRF
- ☐ Include AI-driven attack scenarios, such as a poisoned instruction file, in threat assessments and scenario tests. SEBI advisory circular, 5 May 2026
- ☐ Keep a person in the loop and a way to stop any AI tool, including third-party tools. SEBI AI/ML guidelines, announced on 14 November 2026 and not yet issued
Insurers
- ☐ Map coding assistant controls to the insurer cyber security guidelines. IRDAI Information and Cyber Security Guidelines, 2023
All BFSI entities
- ☐ Report a cyber incident to CERT-In within 6 hours of noticing it. CERT-In Directions, 28 April 2022
- ☐ Keep agent and system logs for 180 days, within India. CERT-In Directions, 28 April 2022
- ☐ Apply reasonable security safeguards to personal data that agents can reach. DPDP Act Section 8(5), DPDP Rules Rule 6
- ☐ Prepare the DPDP breach steps: notice to the Data Protection Board without delay and a detailed report within 72 hours, from 13 May 2027. DPDP Rules Rule 7
For help with the privacy side, see our DPDP consulting service and ISO/IEC 27701 consulting.
Standards and rules referenced
Each layer answers a different question about the same attack.
| Layer | Standard or rule | What it answers for this attack |
|---|---|---|
| Board | IEEE 2863-2026 | Who sets the risk appetite for agents, and who receives red team results |
| AI management system | ISO/IEC 42001:2023 | Inventory, risk assessment, supplier control and logging for AI you use |
| AI security | ISO/IEC 27090:2026 | Prompt injection, supply chain attacks and least privilege for agents |
| Information security | ISO/IEC 27001:2022 | Access, source code, malware, secure development and incident response |
| Privacy | ISO/IEC 27701:2025, GDPR, DPDP Act | Personal data in the blast radius and breach notice |
| EU product law | EU AI Act, EN 18286, MDR, IEC 62304, IEC 81001-5-1 | Duties when the code ends up in a regulated device |
| EU sector security | NIS2 | Management approval, supply chain security and 24 hour early warning |
| India BFSI | RBI directions, FREE-AI, SEBI CSCRF, IRDAI guidelines, CERT-In | Outsourcing, 6 hour reporting, data storage and AI governance |
Clause numbers for ISO/IEC 42001, 27001, 27090 and IEEE 2863 are checked against the published texts. Our ISO/IEC 27090 map links all nine AI attacks to ISO/IEC 42001 controls, and our ISO/IEC 42001 gap assessment shows where your AI management system stands.
We help organisations bring coding assistants into their ISO/IEC 42001 AI management system and their ISO 27001 ISMS, and we red team agent setups. Send us a note and we will be in touch.
Frequently asked questions
Yes. In the November 2026 case, files such as AGENTS.md and CLAUDE.md told the agent to run a setup script and not inspect it. The agent read those files as instructions. This is indirect prompt injection, and it needs no flaw in the model itself.
Yes. An AI management system covers AI the organisation uses as well as AI it builds. Annex A controls on intended use (A.9.4) and suppliers (A.10.3) apply directly to a coding assistant bought as a service.
ISO/IEC 27090:2026 describes prompt injection, including indirect injection through external content, in clause 6.9. Its clause 6.11.3.3 on least privilege for AI systems and agents limits what an injected instruction can do.
Isolate the machine, rotate every token, key and cloud secret it could reach, and check repositories for new commits. In India, report the cyber incident to CERT-In within 6 hours. BFSI entities also report to their regulator.
Sources
- Bitdefender Labs: How AI agents become new targets in GitHub malware campaign
- Cyber Security News: Hackers hijack AI coding agents, 18 November 2026
- VDE: Cybersecurity for medical devices under MDR, IVDR and IEC 81001-5-1
- NIS2 incident reporting timelines
- RBI Master Direction on IT Governance, Risk, Controls and Assurance Practices
- Khaitan & Co: RBI Master Direction on Outsourcing of IT Services
- RBI issues FREE-AI framework
- SEBI advisory on AI-driven vulnerabilities, 5 May 2026
- DSCI: Insight brief on the DPDP Rules
Sandhya Khamesra is the Founder and CEO of Pricoris LLP, a Noida-based consultancy specialising in data privacy, cybersecurity governance and organisational resilience. She brings more than 35 years of experience across information security, privacy and risk management, combining a Chartered Accountant’s discipline with hands-on leadership of ISO implementation programmes.
She has led ISMS, PIMS, BCMS and AI governance programmes in regulated and high-risk environments across more than 35 countries. Her work spans India’s DPDP Act, the GDPR, ISO/IEC 27701 and ISO/IEC 42001, with a consistent focus on making privacy work in day-to-day delivery rather than on paper.
Sandhya is known for translating regulatory and governance requirements into practical, audit-defensible operating practice for global IT services teams. She works regularly with boards, regulators and delivery organisations to establish privacy as a professional discipline, not a legal formality.