risk
Adversarial attacks, data poisoning and prompt injection
Data-poisoning corrupts training data and embeds backdoors; adversarial evasion, prompt injection, and jailbreaks fool deployed models at inference; model extraction steals proprietary weights/logic — enabling harmful or policy-violating outputs.
In catalog since 2026-09-17T22:28:00Z · Last changed 2026-09-17T22:28:00Z (f368a6cce277)
Record JSON · Open in map · Data retrieval guide
Catalog revision: 791ff2dd3a45707290badee660f185e514d15f1cf518908628f425c2f2c56ee4. A connection does not establish full coverage.
Attributes
- category
- ai_governance
- domain
- AI Governance
- Secure Development (SDLC) & Application Security
- Vulnerability & Patch Management
- taxonomy
- nist-ai-rmf-risk
- iso-23894-ai-risk
- inherent_rating
- high
Details
- risk_id
- ai-adversarial-poisoning-attacks
- category
- ai_governance
- likelihood
- medium
- impact
- high
- inherent_rating
- high
- treatment
- mitigate
- taxonomies
- nist-ai-rmf-risk
- iso-23894-ai-risk
Source
No record-specific source URL is provided.
Connections
- UC-VULN-09 — Employ non-persistence and information-resilience techniques mitigates Adversarial attacks, data poisoning and prompt injection
- strength
- related
- rationale
- Sourcing critical data from diverse suppliers and refreshing/purging potentially corrupted data reduces reliance on any single poisoned source.
- UC-AI-21 — Commission independent third-party AI evaluations on a quarterly cadence mitigates Adversarial attacks, data poisoning and prompt injection
- strength
- primary
- rationale
- Quarterly independent adversarial-robustness and jailbreak testing is the assurance that injection defenses actually hold.
- UC-SDLC-05 — Enforce secure coding and input validation standards mitigates Adversarial attacks, data poisoning and prompt injection
- strength
- primary
- rationale
- Rejecting or safely encoding unsafe input at trust boundaries is a first-order defense against prompt-injection/jailbreak inputs.
- UC-AI-18 — Defend AI interfaces against adversarial input, injection, and endpoint abuse mitigates Adversarial attacks, data poisoning and prompt injection
- strength
- primary
- rationale
- Screening prompts, retrieved content, and tool results for injection and jailbreak patterns, plus adversarial-input detection, is the inference-time defense against prompt injection.
- UC-AI-15 — Fulfill general-purpose AI model provider obligations mitigates Adversarial attacks, data poisoning and prompt injection
- strength
- primary
- rationale
- Mandated state-of-the-art adversarial testing plus model and infrastructure cybersecurity directly defend against adversarial attacks and model extraction.
- UC-AI-09 — Govern AI data quality, provenance, and preparation mitigates Adversarial attacks, data poisoning and prompt injection
- strength
- related
- rationale
- Provenance recording and data-selection governance reduce training-data poisoning and embedded backdoors.
- UC-SDLC-04 — Engineer systems with secure architecture and design mitigates Adversarial attacks, data poisoning and prompt injection
- strength
- primary
- rationale
- Engineering solutions to remain robust and resilient against adversarial manipulation (EU AI Act Art.15) directly hardens systems against evasion, poisoning and prompt injection.
- UC-AI-14 — Manage responsible AI with suppliers and customers mitigates Adversarial attacks, data poisoning and prompt injection
- strength
- related
- rationale
- Supplier vetting reduces backdoored or poisoned third-party models and components entering the pipeline.