A defense you haven't tested is a hope, not a control. The final discipline of LLM security is treating your guardrails as a system to be measured, attacked, and monitored continuously — because the threat evolves, your application changes, and a defense that worked last quarter can silently rot. This closing post covers red-teaming your own system, operating it in production, and the honest state of the art.
A defense you haven't tested is a hope, not a control. The final discipline of LLM security is treating your guardrails as a system to be measured, attacked, and monitored continuously — because the threat evolves, your application changes, and a defense that worked last quarter can silently rot. Red-teaming, production monitoring, and the honest state of the art.
"Guardrails" is the umbrella term for the runtime checks that sit around a model and screen what goes in and comes out — content classifiers, moderation models, topic and format validators, PII detectors. They're a real and useful layer, distinct from the architectural defenses, and they come with their own design rules: layer them, fail safe, and never mistake them for a wall.
"Guardrails" is the umbrella term for the runtime checks around a model — content classifiers, moderation models, topic and format validators, PII detectors. They're a real and useful layer, distinct from architectural defenses, with their own design rules: layer them, fail safe, and never mistake them for a wall. This post covers building the layered defense in practice.
Injection defense usually focuses on what goes into the model. But an equally dangerous class of bug lives on the way out: whatever the model produces gets passed to another system — a browser, a shell, a database, another service — that trusts it. If the model can be made to emit a malicious payload, and your code renders or executes it, the injection escapes the model and lands in your infrastructure.
Injection defense usually focuses on what goes into the model, but an equally dangerous class of bug lives on the way out: whatever the model produces gets passed to another system that trusts it. If the model can be made to emit a malicious payload and your code renders or executes it, the injection escapes the model and lands in your infrastructure — XSS, SQLi, SSRF, exfiltration.
Everything before this post raised the probability barrier against injection. This post lowers the impact — and impact is what actually protects you. The load-bearing defense against prompt injection isn't a prompt or a filter; it's an architecture where a fully-hijacked model still can't do anything catastrophic, because it was never granted the power to.
Everything before this raised the probability barrier against injection; this post lowers the impact — and impact is what actually protects you. The load-bearing defense isn't a prompt or a filter but an architecture where a fully-hijacked model still can't do anything catastrophic: least privilege, the confused-deputy trap, human-in-the-loop, and the dual-LLM pattern.
Between filtering the input and re-architecting the system sits a tempting middle ground: make the prompt itself more resistant. Delimiters, spotlighting, instruction placement, and defensive system prompts all raise the cost of an attack. None of them close the hole — because they are all still text in the one channel the attacker also writes to — but used well they meaningfully shift the odds.
Between filtering input and re-architecting the system sits a tempting middle ground: make the prompt itself more resistant. Delimiters, spotlighting, instruction placement, and defensive system prompts all raise the cost of an attack — but none close the hole, because they're all still text in the one channel the attacker also writes to. The value is knowing exactly how much they buy.
The first instinct when facing prompt injection is to inspect the input and block the bad stuff. It's a reasonable layer — but a treacherous one, because it creates a feeling of safety far larger than the protection it provides. This post covers the input-side defenses that are genuinely worth having, and draws a hard line around what they can and cannot do.
The first instinct against injection is to inspect the input and block the bad stuff. It's a reasonable layer but a treacherous one — it creates a feeling of safety far larger than the protection it provides. This post covers the input defenses genuinely worth having, and draws a hard line around what they can't do (starting with: indirect injection bypasses them entirely).
You can't defend against what you can't categorize. Prompt-based attacks come in a few structurally distinct shapes — direct injection, indirect injection through content the model reads, and jailbreaks that target the model's safety training — and each demands a different defense. This post maps the attack surface so the rest of the series can defend it systematically.
You can't defend what you can't categorize. Prompt-based attacks come in structurally distinct shapes — direct injection, indirect injection through content the model reads, and jailbreaks targeting the model's safety training — and each demands a different defense. This post maps the attack surface so the rest of the series can defend it systematically.
Prompt injection is the defining security problem of LLM applications, and — unlike SQL injection, which it superficially resembles — it has no clean fix. The reason is structural: a language model reads instructions and data through the same channel, and cannot reliably tell which is which. This opening post explains why that makes injection fundamentally hard, and reframes the goal from "prevent it" to "contain the blast radius."
Prompt injection is the defining security problem of LLM applications, and unlike SQL injection it has no clean fix — because a model reads instructions and data through the same channel and can't reliably tell them apart. This opening post explains why that makes injection fundamentally hard, and reframes the goal from prevent to contain the blast radius.
The frontier is genuinely exciting and genuinely oversold, and telling the difference matters — so this closing post is an honest accounting of what self-evolving agents cannot yet do, what can go wrong, and what remains unsolved.
The frontier is genuinely exciting and genuinely oversold. An honest accounting of what self-evolving agents cannot yet do, what can go wrong, and what remains unsolved.
Every frontier method is a search, and a search is only as good as the evaluator that ranks its candidates — so at the frontier, evaluation stops being a measurement and becomes the single most dangerous component in the system.
Every frontier method is a search, and a search is only as good as the evaluator that ranks its candidates — so at the frontier, evaluation becomes the single most dangerous component in the system.
An MCP server can run code and see context on the model's behalf, which makes it powerful and dangerous in equal measure — this is how to deploy one without handing attackers the keys.
An MCP server can run code and see context on the model's behalf. Authentication, prompt-injection and tool-poisoning risks, human-in-the-loop, sandboxing, and a production checklist.
AI systems add attack surface that traditional security does not cover — the model, its prompts, its retrieved context, and its tools are all attackable — and the only way to know you're defended is to threat-model the whole surface and prove it with red-teaming.
AI adds attack surface conventional security misses — model, prompts, retrieved context, and tools are all attackable. Phase 7: threat-model the whole surface and prove it with red-teaming.
A system that changes itself can improve itself right off a cliff, so the evaluation and guardrails are not an afterthought to self-evolving agents — they are the thing that makes them safe to run at all.
A system that changes itself can improve right off a cliff. Measuring evolution honestly, reward hacking, drift and collapse, and the guardrails that keep it safe.
A single pre-launch red-team decays the moment your model, prompt, or tools change — turning adversarial testing into a sustained program is what keeps an AI system safe past day one.
The capstone: making red-teaming a sustained program — the remediation loop where findings become regression tests, blending automated/manual/external modes, rules of engagement, ship-blocking severity thresholds, governance evidence, and culture.
The series finale: how to proactively find AI security failures before attackers do — turning injection, leakage, and excessive-agency risks into a repeatable adversarial test suite that runs in CI, measuring attack success honestly, and standing up incident response for the day a control fails.
Find AI security failures before attackers do: adversarial testing of the whole system, a test taxonomy mapped to the series, automated tooling (PyRIT, garak), a CI security-regression suite that fails the build, honest attack-success-rate measurement, and AI incident response.
Turning red-team attacks into metrics you can act on and track over time — attack success rate, coverage, severity, and trend — plus the honest limits of what any of those numbers can tell you.
Turning attacks into metrics: attack success rate and why it's subtle, scoring success (rule/classifier/LLM-judge with its biases), coverage across the taxonomy, severity weighting, tracking trends per model/prompt version, and honest reporting of residual risk.
Part seven of the AI Security Engineering series: DevSecOps for AI systems — securing the secrets, network, supply chain, prompts, and CI/CD gates that surround the model, so a hardened model doesn't sit inside a soft pipeline.
DevSecOps for AI: secrets in a manager not code, least-privilege runtime identities (no ambient prod creds for agents), egress control, rate/spend limits against model DoS, supply-chain verification in CI, prompts-as-code, and a security-eval gate that fails the build.
Scaling red-teaming beyond manual probing — the building blocks of an automated harness (seed library, mutation, orchestrator, scorer), LLM-driven adaptive attackers, the real tools by role (PyRIT, garak, promptfoo, Giskard), and wiring it all into CI as a repeatable gate.
Scaling red-teaming: the harness building blocks (attack seeds, mutation, orchestrator, scorer), adaptive LLM-driven attackers, the real tools by role (PyRIT, garak, promptfoo, Giskard), and integrating an automated red-team gate into CI.
The defensive layer that screens what goes into a model and what comes out — input rails, output rails, topical rails, and groundedness checks — plus the real tooling ecosystem and a vendor-neutral Python pipeline that wraps a model call and knows how to refuse.
The defensive layer that screens inputs and outputs: input/output/topical/groundedness rails, the real ecosystem (Llama Guard, Granite Guardian, NeMo Guardrails, Guardrails AI, Presidio, hosted moderation), and a vendor-neutral guardrail pipeline — with honest false-positive/negative trade-offs.
Why agents and retrieval turn a prompt injection into real-world action, how to red-team the highest-risk AI surface with benign canaries, and the least-privilege controls that shrink an attacker's blast radius.
The highest-risk modern surface: indirect injection via RAG/tools, tool abuse and excessive-agency exploitation, memory poisoning, multi-step attacks, and data-exfiltration channels — with a canary methodology and least privilege as the primary control.
Part five of the AI Security Engineering series: the two OWASP LLM risks that live in the plumbing around the model — the documents your agent retrieves and the models, datasets, and dependencies it is built from — and the Python patterns that treat both as untrusted until proven otherwise.
Securing the components around the model: RAG as an injection and poisoning vector (treat retrieved content as hostile, enforce provenance and per-user authz) and supply-chain risk (safetensors over pickle, model provenance, pinned and vetted dependencies and tools).
A defender's tour of the attacks that target the model and its data — prompt and context extraction, training-data memorization, membership inference and model inversion, model stealing, poisoning and backdoors, and evasion — with what a red-teamer tests and what actually stops each one.
Attacks on the model and its data: system-prompt and training-data extraction, membership inference and model inversion, model stealing, and poisoning/backdoors — what to test and defend, and the honest risk difference between using hosted models and training your own.
Two tightly-linked OWASP LLM risks that turn a clever prompt injection into real-world damage — and the Python patterns that shrink the blast radius: treat model output as untrusted input, and give agents the least agency they can get away with.
Two OWASP risks that turn an injection into damage: insecure output handling (model output is untrusted input — never eval/shell/SQL it unescaped) and excessive agency (least-privilege tools, allow-lists, human approval for irreversible actions, audit logs).
A defender's field guide to the injection and jailbreak techniques a red-teamer probes for — the taxonomy, why each one works, and how to turn it into a re-runnable test suite that maps every passed test to a concrete fix.
A deeper, test-focused look at injection and jailbreak families — direct vs indirect (the RAG/agent threat), role-play, obfuscation, many-shot, multi-turn, cross-lingual — plus a red-team methodology: build a probe suite, mutate, test input and retrieval paths, measure, re-test.
Part three of the AI Security Engineering series: protecting the data that flows through an LLM system — how sensitive information leaks out of prompts, logs, and retrieval, and the engineering controls (redaction, data minimization, per-user retrieval authz, residency choices) that actually stop it.
Protecting data in LLM systems: sensitive-information disclosure, PII in prompts and logs (your observability can be the leak), redaction with Presidio, data residency, and per-user access control on the retrieval layer so RAG doesn't leak across tenants.
Before you attack an AI system you need a map of it: the components an adversary can influence, the trust boundaries between them, and a taxonomy that sorts attacks by goal and stage so your red-teaming is systematic instead of a grab-bag of the attacks that happen to trend that week.
Mapping the AI attack surface so red-teaming is systematic: threat-modeling the components (model, prompts, training/RAG data, tools, guardrails) and an attack taxonomy by goal (integrity/availability/privacy/abuse) and stage (training vs inference), aligned to NIST AI 100-2 and ATLAS.
Why the #1 risk on the OWASP LLM Top 10 has no clean fix — the model can't tell your instructions apart from the text it reads — and the defense-in-depth pattern that actually shrinks the blast radius.
The #1 LLM risk: direct vs indirect prompt injection (the latter is the real threat for RAG and agents), jailbreak families, and honestly-rated defense-in-depth — delimiting untrusted data, least-privilege tools, human-in-the-loop, and detection's limits.
The opening post of a hands-on series for builders: what it means to systematically stress-test an AI system — model, prompts, retrieval, tools, and guardrails — to surface its failures before adversaries or ordinary users do, how that differs from pentesting and robustness testing, and the frameworks and ethics that keep the work rigorous.
The opener to a defensive AI red-teaming series: stress-testing AI systems (models + prompts + RAG + tools + guardrails) to find failures before adversaries do — how it differs from pentesting, the frameworks (OWASP GenAI, MITRE ATLAS, NIST AI 100-2), and ethical rules of engagement.
Why LLM and agent applications open a genuinely new attack surface, the mental models to reason about it (OWASP Top 10 for LLM Applications, MITRE ATLAS, NIST AI RMF), and how to threat-model an AI system before you write a line of defensive code.
The opener to a series on securing LLM and agent systems: why AI apps have a new attack surface (instructions and data share one channel, tools grant real power), the OWASP Top 10 for LLM Applications, MITRE ATLAS, and how to threat-model an AI system.