Threat Intelligence
Recent vulnerability activity pulled directly from NVD and CISA at page load, alongside the frameworks we ground our own testing and controls in: OWASP's LLM and Agentic Top 10s, MITRE ATLAS, NIST's AI Risk Management Framework, and ISO/IEC 42001.
Recent CVEs — AI & LLM keyword match
Recently added — CISA Known Exploited Vulnerabilities
OWASP Top 10 for LLM Applications — 2025
genai.owasp.org →Prompt Injection
Crafted input overrides system instructions, directly or via hidden content the model processes.
Sensitive Info Disclosure
Training data, secrets, or context bleed into outputs, logs, or traces.
Supply Chain
Compromised models, datasets, libraries, or hosting providers in the pipeline.
Data & Model Poisoning
Malicious training, fine-tuning, or retrieval data shapes model behavior.
Improper Output Handling
Downstream systems trust model output without validation or sanitization.
Excessive Agency
Excessive functionality, permissions, or autonomy granted to a model or agent.
System Prompt Leakage
Instructions, keys, or internal logic embedded in the system prompt get exposed.
Vector & Embedding Weaknesses
Poisoned, unauthorized, or manipulated access to RAG vector stores.
Misinformation
Confident, plausible, and false output — hallucination treated as fact.
Unbounded Consumption
Resource-heavy prompts or loops drive cost, denial-of-service, or model theft.
OWASP Top 10 for Agentic Applications — 2026
genai.owasp.org →Agent Goal Hijack
Malicious content alters an agent's objective or reasoning path mid-task.
Tool Misuse & Exploitation
Legitimate tools get bent into destructive or unintended actions.
Identity & Privilege Abuse
Agents operate with broader permissions than their task requires.
Agentic Supply Chain Compromise
Compromised skills, plugins, or agent-to-agent dependencies.
Unexpected Code Execution
Agent-generated or agent-triggered code runs beyond its intended sandbox.
Memory & Context Poisoning
Poisoned long-term memory or embeddings reshape future decisions.
Insecure Inter-Agent Comms
Spoofed or tampered messages between agents in a multi-agent system.
Cascading Failures
Small errors propagate and amplify across chained agents and workflows.
Human-Agent Trust Exploitation
Confident agent output leads a human to approve a harmful action.
Rogue Agents
A compromised or misaligned agent acts harmfully while appearing legitimate.
Frameworks we ground our methodology in
MITRE ATLAS
A living matrix of real-world adversary tactics and techniques against AI/ML systems — reconnaissance, ML model access, staging, exfiltration, and impact — built on the same structure as ATT&CK, with case studies of actual incidents.
atlas.mitre.org →NIST AI Risk Management Framework
Four core functions — Govern, Map, Measure, Manage — for identifying and reducing AI risk across an organization. We use it as the shared vocabulary for translating red-team findings into governance action.
nist.gov →ISO/IEC 42001
The first international standard for an AI management system — governance, risk, and lifecycle controls for organizations that build or deploy AI. Referenced as a target for clients pursuing formal certification.
iso.org →On sourcing: the two panels above call the NVD and CISA APIs directly from your browser at page load — we don't proxy, cache, or store what they return. The CVE search is keyword-matched against NVD for "large language model" and "artificial intelligence," which surfaces relevant results but isn't a guarantee of completeness — always cross-check against the source for anything decision-critical. The CISA KEV catalog isn't AI-specific; it's shown because it's one of the highest-signal "what's actively being exploited right now" feeds available publicly. The OWASP, MITRE, NIST, and ISO summaries above are maintained by us and refreshed periodically against the source publications; they are not live API calls, since none of those bodies publish one for this content — click through to each source for the authoritative, current version.