All articles
AI attack techniques against deployed LLMs and agents: jailbreaks, prompt injection, model extraction, evasion, and data poisoning. Every technique is documented from published research, disclosed incidents, and vendor documentation, with the source named so you can check it.
-
Prompt Jailbreak Examples: 7 Patterns Explained
Seven prompt jailbreak examples from published research, explained by mechanism: what each does to the model and how defenders catch it.
-
RAG Poisoning: Corrupting Retrieval and Embeddings
RAG poisoning corrupts the retrieval corpus or embedding store so a model pulls attacker-controlled context at inference, without touching its weights.
-
How Indirect Prompt Injection Works
How indirect prompt injection works: attacker instructions hidden in web pages, emails and RAG documents, the attack surface, and what defenders can do.
-
LLM Jailbreak Defenses: Why Static Filters Fail
Why LLM jailbreak defenses that score well on static benchmarks collapse against adaptive attackers, and what a robustness evaluation has to measure.
-
LLM Jailbreak Examples: 10 Documented Patterns
Ten LLM jailbreak examples drawn from published research, with the reported success rates, the mechanism behind each, and the signals that detect them.
-
Adversarial Examples Explained: How Pixels Fool Models
A plain-language walkthrough of adversarial examples: what they are, why neural networks fall for them, how FGSM works, and what defenders can do.
-
OWASP Top 10 LLM Explained: Every Entry and What to Fix
The OWASP Top 10 for LLM Applications 2025 explained: every entry, what it means in production, what the revision changed, and which fixes return most.
-
Evasion Attacks in Production ML: Malware, Spam, and Fraud
Deployed ML classifiers in malware, spam, and fraud detection face evasion attacks where the attacker has a clear payoff.
-
Web-Scale Dataset Poisoning: Split-View and Frontrunning
Split-view attacks exploit mutable web content, while frontrunning targets timed snapshots to poison datasets without training pipeline access.
-
Adversarial Examples Against Vision Models in 2026
Where physical-world adversarial patches and digital attacks stand against modern vision models, what has been hardened, and what red teams should test.
-
How GCG Works: Greedy Coordinate Gradient, Explained
A mechanics-first look at Greedy Coordinate Gradient (GCG): how the algorithm optimizes tokens over discrete space, why suffixes transfer, and its limits.
-
Indirect Prompt Injection in RAG Pipelines
How attackers embed malicious instructions in documents that get retrieved into LLM context — and why RAG makes prompt injection a supply-chain problem.
-
Jailbreaking Multimodal Models: Visual Prompt Injection
How attackers use images, typography, and adversarial perturbations to bypass safety guardrails in multimodal models, and why vision widens the surface.
-
LLM Context Window Poisoning
How attackers plant persistent malicious instructions in memory and context, survive summarization, and carry influence across separate LLM sessions.
-
Many-Shot Jailbreaking: What the Research Shows
How hundreds of synthetic compliance examples in a long-context prompt erode safety training, what the published results show, and why patching is hard.
-
Model Extraction via Black-Box Query Attacks
How attackers reconstruct model behavior and decision boundaries through query-only access, the economics of the attack, and what stolen models enable.
-
Model Inversion and Membership Inference in LLMs
How membership inference reveals whether specific data trained a model, and how inversion reconstructs private text from gradients and model outputs.
-
Supply Chain Attacks on AI Models: Poisoning and Backdoors
How attackers compromise models before production through dataset poisoning, malicious fine-tuning, and serialization exploits, and how to detect it.
-
Tool-Call Hijacking in Agentic Systems
How attackers exploit the gap between LLM reasoning and tool execution to trigger unauthorized tool calls, and why confirmation gaps are the root cause.
-
Building a CI Gate for Prompt Injection Regression
A practical CI gate that catches prompt injection regressions before release, so a routine prompt change cannot silently weaken your guardrails.