← All posts

Problem-Aware

249 posts on problem-aware.

LLM Request Smuggling with Content-Length and Transfer-Encoding

LLM request smuggling begins when two HTTP recipients disagree about where one request ends and the next begins. Conflicting Content-Length and Transfer-Encoding fields are the classic fault line. In an AI gateway chain, the disagreement can place hidden bytes beyond the inspected prompt and ahead of the next authenticated request. This guide covers strict rejection, canonical reserialization, connection handling, protocol translation, and negative tests.

ai-securityllm-securitycybersecurityzero-trustinline-enforcement
Read post →

OWASP AIVSS: Building an Evidence-Backed Agentic Vulnerability Score

OWASP AIVSS extends vulnerability assessment with AI-specific, temporal, environmental, and modified metrics. A useful assessment preserves the evidence behind every selected value, names the deployment context, and separates routed observations from model, training, infrastructure, and governance evidence. This guide turns the score into a reviewable security record.

ai-securityagentic-aillm-securityai-governanceauditarchitecture
Read post →

Vercel AI SDK Prompt Injection: Control Every Agent Step

Vercel AI SDK agents can add tool results and new context before each model step, so the first input check sees only part of the eventual prompt. Limit active tools, validate tool inputs, require approval for consequential actions, inspect each assembled model request, and keep application-side execution controls separate from the HTTP LLM boundary.

ai-securityagentic-aiprompt-injectionllm-securitypolicy-enforcement
Read post →

Tokenizer Attacks Exploit the Gap Between Text and Tokens

Tokenizer attacks craft text whose token representation degrades model understanding or evades text-level assumptions. Defenses need Unicode normalization, raw and normalized evidence, model-specific token-budget tests, route policy, and response checks rather than one universal character filter.

ai-securityllm-securityprompt-injectionpolicy-enforcementarchitecture
Read post →

Sponge Examples Make Worst-Case LLM Cost a Security Requirement

Sponge examples are crafted inputs that maximize neural-network energy use or latency. LLM operators should test worst-case inputs, enforce route-specific size and time budgets, cap concurrency, preserve request evidence, and stop repeated expensive payloads before inference.

ai-securityllm-securitypolicy-enforcementarchitecturecybersecurity
Read post →

Reasoning Model Jailbreaks Scale the Adversary, Too

Large reasoning models can plan and adapt multi-turn jailbreaks against other models. Enterprises need separate identities for attacking and target models, bounded routes, conversation-level inspection, response controls, and audit records that preserve the full adversarial chain.

ai-securityllm-securityprompt-injectionagentic-aipolicy-enforcement
Read post →

QR Code Prompt Injection Hides Instructions in the Image Path

A multimodal agent can decode a QR code inside an image and treat the recovered text as instructions. Control the image path with provenance, decoder output inspection, task-scoped tool permissions, model-route policy, and records that connect the original image to the model call.

ai-securityprompt-injectionagentic-aillm-securitypolicy-enforcement
Read post →

Policy Puppetry Attacks Turn Policy Syntax Against Model Guardrails

Policy Puppetry wraps harmful intent in policy-like syntax and role-play so frontier models treat the payload as authoritative structure. Defenses need adversarial testing, request classification, identity-bound policy, constrained routes, and response inspection outside the model.

ai-securityllm-securityprompt-injectionpolicy-enforcementcybersecurity
Read post →

PDF Hidden-Text Prompt Injection Survives the Human Review

A PDF can look ordinary on screen while its extracted text contains instructions for an LLM. Secure ingestion preserves provenance, compares visible and extracted content, treats document text as data, and inspects the assembled model request before transmission.

ai-securityprompt-injectionllm-securitydata-loss-preventionpolicy-enforcement
Read post →

OverThink Slowdown Attacks Turn Context Into a Reasoning Bill

An OverThink attack places decoy reasoning problems inside external context so a reasoning model spends extra compute before returning an otherwise plausible answer. The practical controls are context provenance, input budgets, route limits, timeouts, and per-request cost evidence.

ai-securityllm-securityprompt-injectionpolicy-enforcementarchitecture
Read post →

Bad Likert Judge Jailbreak Turns Evaluation Into the Attack

Bad Likert Judge asks an LLM to score harmful responses on a Likert scale, then generate examples that match the ratings. Unit 42 tested 1,440 cases across six anonymized models and found a large increase in attack success. The deployment response needs prompt and response inspection, scoped authorization, and evidence at the routed HTTP boundary.

llm-securityai-securityprompt-injectionllmpolicy-enforcement
Read post →

Training Data Extraction Attacks Leave a Query Trail

A training data extraction attack uses model queries to recover memorized training content. NIST separates extraction from membership inference and model extraction, while the original GPT-2 experiment shows how black-box generation and ranking can recover verbatim sequences. This article assigns training, deployment, and incident controls to their proper owners and defines the evidence available at a routed HTTP model boundary.

ai-securityllm-securitydata-loss-preventionidentity-and-authorizationpolicy-enforcementforensic-audit
Read post →