Decepticon wraps web crawl output into LLM messages without neutralizing ChatML special-token literals.
Under BYOK, most self-hosted inference stacks do not filter those literals by default.
Read AdvisorySecurity Blog
Expert insights on AI security, cybersecurity best practices, and cutting-edge security tools
Latest Blog Articles
resk-logits Benchmark: Token-Level Filtering Latency
Measured CPU overhead from 100 to 10,000 patterns, with the exact command, raw JSON and the caveats we refuse to hide.
Read ArticleLLM Guard Is Archived: Migration Guide and Alternatives
What the July 2026 archive actually changes, the four realistic options, and a one-line scanner migration path with the compatibility shim.
Read ArticleLeast Privilege for AI Agents: From Entitlement Mapping to Enforcement
Map every agent entitlement, derive capability masks by intersection, enforce at generation and tool-call time, isolate secrets and audit drift.
Read ArticleMCP Security Mechanisms and Layers: A Technical Deep Dive
Six security layers for the Model Context Protocol — transport, authentication, manifest integrity, tool authorization, output handling and audit.
Read ArticleEnterprise LLM Governance: From Policy Document to Enforced Control
Ownership, access control, PII handling, audit logs, model registry and policy-as-code — the parts of an AI governance policy that can actually be tested.
Read ArticleLLM Security Monitoring & Observability: What to Log and What to Detect
Five telemetry planes, eight high-signal detections, privacy-safe logging and a 30-day rollout for LLM security monitoring.
Read ArticleAI Penetration Testing: A Framework for LLMs and Agents
A repeatable six-phase method: scope, threat model, versioned corpus, agent and MCP testing, CI automation and remediation-focused reporting.
Read ArticleTypeScript Secure Coding for AI Applications
Bundled API keys, unvalidated tool arguments, prototype pollution, SSRF and XSS — the TypeScript-specific mistakes in LLM apps, with fixes.
Read ArticlePrompt Leaking: How System Prompts Leak and How to Stop It
How attackers extract system prompts and sensitive context — refusal loops, encoding tricks, context switching — and the defenses that work: canary tokens, output filtering, generation-time enforcement.
Read ArticleReskSafety: The Deployable Open-Source LLM Firewall
Meet ReskSafety: an OpenAI-compatible LLM firewall with RBAC capability bitmasks, editable filtering policies, tool-call authorization, multi-provider routing, and a full admin console. Try it live at demo.resk.fr.
Read ArticleSelf-Host an LLM Firewall with Docker in Minutes
A step-by-step guide to deploying ReskSafety — an open-source, OpenAI-compatible LLM firewall — on your own infrastructure with Docker Compose, from clone to enforced policies.
Read ArticleCapability Bitmasks: Real RBAC for LLM Applications
Why LLM apps need per-user permission enforcement, and how 64-bit capability bitmasks enforce least privilege at the token level — so unauthorized content is never generated.
Read ArticleLLM Guardrails vs LLM Firewalls: Which One Do You Need?
Guardrails shape behavior; firewalls enforce boundaries. Where each fails, and how production systems combine both.
Read ArticleTool Permissions for AI Agents: Least Privilege in Practice
Agents with tools are privileged insiders. Capability bitmasks, generation-time gating, and post-call verification — applied step by step.
Read ArticleShadow Banning: Filter LLM Output Without Breaking the User Experience
Hard blocks crash conversations. Shadow banning makes harmful output statistically vanish — silent, invisible, jailbreak-resistant.
Read Article8 Open Source Libraries to Secure Your LLM Applications
Discover the RESK open-source ecosystem: 8 free libraries to secure your LLM applications across Python, TypeScript, and Bun. Free on GitHub, PyPI, and npm.
Read ArticleAI Red Teaming: The Complete Guide to LLM Security Testing
Master AI red teaming: adversarial testing techniques, prompt injection simulations, jailbreak detection, and how to harden LLM applications against real-world attacks.
Read ArticleMCP Security: Protecting AI Agents from Model Context Protocol Threats
Understand the security risks of the Model Context Protocol and how resk-mcp protects AI agents from tool-based attacks and prompt injection through tool outputs.
Read ArticleEnd-to-End Encryption for AI Conversations: The ReskCrypt Approach
Protect sensitive AI conversations with end-to-end encryption. Discover ReskCrypt's cryptographic approach to LLM data privacy and policy-based access.
Read ArticleSecuring AI Applications: Architecture Patterns That Work
Where to place each control so it actually enforces: embedded libraries, an enforcement gateway in the request path, and action authorization for agents — with a worked request flow.
Read ArticlePrompt Injection Attacks: Case Studies and Lessons Learned
From Bing "Sydney" to EchoLeak and the Copilot Chat RCE: how five real prompt injection attacks worked, what each one teaches, and the defenses that map to each failure.
Read ArticleZero Trust AI Security Architecture
Implementing Zero Trust principles in AI systems. Research on trust boundaries, continuous verification, and securing AI workloads in modern enterprise environments.
Read ArticleEnterprise AI Governance Framework
Establishing robust governance frameworks for enterprise AI deployments. Research on risk management, compliance requirements, and organizational best practices.
Read ArticleAI Security Best Practices: An Implementation Checklist
Best practices only work when they are sequenced. The production checklist for LLM deployments: identity, input, context, output, actions, monitoring, red teaming.
Read ArticleLLM Access Controls & Monitoring: Best Practices
Per-team RBAC, token-level enforcement, audit logs and monitoring for enterprise LLM applications — with a deployment checklist.
Read ArticleImplementing RESK-LLM Security Framework
Technical guide for implementing the RESK-LLM security framework. Step-by-step instructions and configuration examples for secure LLM deployments.
Read ArticleOur Security Focus
Join Our Security Community
Stay updated with the latest security insights and contribute to the advancement of AI security.
Contact Our Team