Prompt Injection
Definition
An attacker hides instructions inside user input or external content (a document, webpage, email) that hijacks the AI model into ignoring its original instructions and doing something else instead.
Our Defense
Input/output segregation, instruction-hierarchy enforcement, and content-source tagging so the model can tell "trusted instruction" from "untrusted data" apart.
Jailbreaking & Guardrail Evasion
Definition
Techniques used to talk a model out of its safety training — role-play framing, encoding tricks, or multi-turn manipulation designed to get restricted content or behavior out of the system.
Our Defense
Layered guardrails outside the model itself (not just relying on the model's own training), adversarial red-teaming, and continuous prompt-pattern monitoring.
Data Poisoning
Definition
Corrupting the data a model is trained or fine-tuned on so it learns a hidden bias, backdoor, or vulnerability that surfaces later — often invisibly, until triggered.
Our Defense
Provenance tracking on every training and fine-tuning dataset, statistical anomaly detection, and staged validation before any dataset touches a production model.
Model Theft & Extraction
Definition
Systematically querying a deployed model to reverse-engineer its weights, architecture, or training data — effectively stealing years of R&D through the API.
Our Defense
Rate limiting, query-pattern fingerprinting, output watermarking, and legal/contractual controls layered on top of the technical ones.
Sensitive Information Disclosure
Definition
A model unintentionally reveals confidential training data, system prompts, API keys, or personal data embedded in what it learned or was given as context.
Our Defense
Output filtering, PII/secret scanning on both input and output, and strict context-window hygiene so sensitive data never enters the prompt in the first place.
Insecure Output Handling
Definition
Treating a model's output as automatically safe and feeding it straight into a downstream system (a database query, a shell command, a rendered webpage) without validation.
Our Defense
Treat every model output as untrusted user input — the same sanitization and validation discipline you'd apply to a public web form.
RAG & Supply-Chain Attacks
Definition
Poisoning the documents a retrieval-augmented system pulls from, or compromising a third-party model, plugin, or dependency the AI system relies on.
Our Defense
Vetted and access-controlled knowledge sources, integrity checks on retrieved content, and a software bill of materials for every model and plugin in the stack.
Excessive Agency
Definition
Giving an agentic AI system more autonomy, tool access, or permissions than the task actually requires — turning a small manipulation into a large real-world consequence.
Our Defense
Least-privilege tool access, human-in-the-loop checkpoints for consequential actions, and hard scoping of what any agent is allowed to touch.
Model Denial of Service
Definition
Deliberately crafted inputs that force a model to consume excessive compute — long-running generations, resource-exhausting prompts, or repeated expensive queries — degrading service for everyone else.
Our Defense
Query complexity limits, per-user throttling, and cost/usage anomaly alerts tied directly to infrastructure monitoring.
Adversarial Examples & Evasion
Definition
Small, often imperceptible changes to an input (an image, a piece of text) that are engineered to make a model misclassify or misbehave, while looking normal to a human.
Our Defense
Adversarial training, robustness testing against known attack libraries, and ensemble/consensus checks for high-stakes classification decisions.
Membership Inference & Model Inversion
Definition
Privacy attacks that determine whether a specific record was in a model's training data, or reconstruct approximations of that training data from the model's behavior.
Our Defense
Differential privacy techniques, training-data minimization, and privacy-impact testing before a model is cleared for deployment.
Deepfake & Synthetic Media Threats
Definition
AI-generated audio, video, or images used for fraud, impersonation, or disinformation — including voice-cloned executives used in social-engineering attacks.
Our Defense
Detection tooling, verification protocols for high-risk requests (payment changes, credential resets), and staff awareness training built specifically around this threat.
aisecuritysimplified.ie™