How can employees' AI accounts and token budgets become an arsenal against the company and its supply chain?
When a company opens generative AI to staff, it usually issues API keys, an enterprise account or a monthly budget. These have three properties: they convert directly into compute, the usage is recorded under the company's name, and they are easy to copy. Real cases exist. On 27 February 2025 Microsoft's Digital Crimes Unit reported that the Storm-2139 network exploited exposed customer credentials scraped from public sources to access generative AI services without authorisation, altered the services' capabilities and resold access to other malicious actors to produce harmful content.
The impact on a company has at least four layers:
- Cost: the budget is drained; OWASP lists this risk as LLM10 Unbounded Consumption.
- Liability and reputation: abusive content comes from the company's account, the provider may suspend it, and the company has to prove its innocence.
- Attacks on yourself: a stolen account or an insider uses the company's AI to write phishing aimed at colleagues or to reconnoitre internal systems.
- Attacks on the supply chain: an AI agent with tool permissions, once hit by prompt injection, can send malicious content to suppliers and customers from your trusted domain.
AI usage boundaries are therefore not just about what may be asked of AI; AI access has to be managed as a high-value credential.
Which misuse and attack scenarios belong in the AI risk assessment?
| Scenario | How it happens | Consequence | Reference |
|---|---|---|---|
| API key leakage | Keys in code, configuration files, shared documents or chat groups | Outsiders do anything on the company's budget | Microsoft Storm-2139 case; OWASP LLM10 |
| Shared accounts, no accountability | One key shared by a whole department | No way to tell who did it or to disable one person | ISO/IEC 42001 Annex A.9 |
| Insider misuse | Company budget used to write phishing or malware or to scrape at scale | Legal liability, service suspension | NIST AI 100-2 misuse |
| Prompt injection into AI agents | The agent reads a malicious mail or page and follows its instructions | External mail, downloads, code execution | NIST indirect prompt injection; OWASP LLM01, LLM06 |
| Budget exhaustion, outage | Runaway loops or abuse | Cost spike, normal work disrupted | OWASP LLM10 |
| No revocation on leaving or transfer | Accounts and keys stay valid | Former staff can still use company AI and data | ISO/IEC 42001 Annex A.9 |
| AI misuse on the supplier side | Contractors process your data or contact others on your behalf with their own AI | Your data and name misused | ISO/IEC 42001 Annex A.10 |
How do you set the usage boundary: what is allowed, what is not, and who approves?
ISO/IEC 42001 Annex A.9 asks the organisation to define processes for responsible use of AI systems (A.9.2), objectives for responsible use (A.9.3) and to ensure AI systems are used according to their intended use (A.9.4). For staff use of AI we recommend three written lists:
- Approved uses: for each use, the models or services allowed, the data classification that may be entered, the tools that may be called, and an owner. Anything not on the list needs an application first.
- Prohibited uses: creating or modifying malware, producing phishing or social engineering content, scanning or scraping other parties' systems without authorisation, circumventing provider safeguards, entering data above the approved classification. Keep the list consistent with each AI provider's usage policy.
- AI agent permissions: what each agent may read and write and whether it may send externally or execute code; high-risk actions always need human confirmation.
AI Act Article 4, as amended by Regulation (EU) 2026/1744, requires providers and deployers of AI systems to take measures to support the AI literacy of their staff. Training staff on these three lists and keeping the records is the most direct way to do it.
How do you control keys, budgets and AI agents technically, and what evidence should you keep?
The controls are our practical recommendations; the basis column follows the standards and legal texts.
| Basis | Control | Evidence |
|---|---|---|
| ISO/IEC 42001 A.9.2, A.9.4 | All AI access through an enterprise AI gateway with single sign-on identifying each person; no long-lived shared keys | Gateway configuration, account list |
| OWASP LLM10 | Budget caps per person and purpose, spend alerts and anomaly detection (heavy use at night, sudden model switches, bursts of similar requests) | Budget settings, alert records |
| Key leakage protection | Secret scanning of code and documents; short-lived, rotatable keys held in a key management service | Scan reports, rotation records |
| OWASP LLM06; NIST indirect prompt injection | Least privilege for agents, external content separated from instructions, human confirmation for external sending and code execution | Permission list, confirmation records |
| ISO/IEC 42001 A.9.3 | Periodic review of usage and purpose records, sampling for use outside approved purposes | Review records |
| Staff changes | AI accounts and keys revoked on the day of leaving or transfer | Revocation records |
| Incident response | On key leakage or misuse: revoke, notify the provider, assess impact on suppliers and customers | Incident records |
How do you manage the supply chain side?
ISO/IEC 42001 Annex A.10 covers third-party and customer relationships: allocating responsibilities between the organisation, its suppliers and customers (A.10.2), establishing a supplier process (A.10.3) and considering customer expectations and needs (A.10.4). For AI misuse the supply chain has to be managed in two directions:
- Your suppliers: when contractors use AI on your data or code, or contact your customers on your behalf, the contract states which AI services may be used, forbids training on your data and sets deadlines for reporting misuse and leaks.
- Your responsibility to suppliers: if your AI accounts or agents are used to attack a supplier, you must be able to disable them at once, notify the supplier and hand over records for the investigation.
Writing both directions into supplier management and incident response keeps AI misuse from turning into mutual blame between you and your suppliers.
How do IT, security, legal and HR divide the work, which part do Winners take, and where do you start?
- AI governance lead: maintains the three lists, approves new uses and chairs periodic reviews.
- IT and security: build the AI gateway, budgets and alerts, secret scanning and agent permissions.
- Legal and procurement: AI clauses in provider and supplier contracts.
- HR: onboarding, training and exit procedures.
Winners Consulting implements ISO/IEC 42001: AI system inventory, usage boundaries and the three lists, risk and impact assessment, supplier clauses and evidence planning, sharing one evidence base with an existing ISO 27001 system. Ahura (POC version) is fully on-premises, keeping approval and review records in the company's own environment.
Start with three things: inventory every AI account and key (including those scattered in code), replace shared keys with personal identities through the gateway, and publish the prohibited-use list first.
AI attack defence for other services
Official sources
- ISO/IEC 42001:2023 — Artificial intelligence management system — ISO
- Regulation (EU) 2024/1689 (Artificial Intelligence Act) — EUR-Lex
- Regulation (EU) 2026/1744 (Digital Omnibus on AI) — EUR-Lex
- OWASP Top 10 for LLM Applications 2025 — OWASP GenAI Security Project
- NIST AI 100-2 E2025, Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations
- Microsoft On the Issues — Disrupting a global cybercrime network abusing generative AI (27 Feb 2025)
