Adversarial ML and Prompt Attacks
Universal Adversarial Suffixes Against Tool-Calling LLMs
Adversarial suffixes hijack tool-calling models to execute unauthorized actions on live systems.
Priya Subramaniam
· 9 min read
ThreatIntel Review
Inter-Agent Trust Boundaries in Multi-Agent Pipeline Architectures
Verifying content, identity, intent, and state at each agent handoff prevents cascade attacks.
Desmond Okwuosa
Goal Hijacking and Memory Poisoning Mechanics in Production Agents
Attackers redirect agent objectives and corrupt memory across sessions without touching code.
Desmond Okwuosa
Attacker Reconnaissance of AI Agent Tool Manifests and Memory Architecture
Attackers exploit tool manifests and memory to map agent capabilities and predict future actions.
Priya Subramaniam
Prompt Injection vs Jailbreak Attacks on AI Agents
Two attack vectors targeting AI agents require separate defenses, not unified guardrails.
Priya Subramaniam
AI Gateway Guardrail Coverage Gaps Across LiteLLM and Kong
Gateway guardrails inspect API boundaries but cannot enforce intent across agent tool chains.
Priya Subramaniam
Output Filtering Gaps in Multi-Step Agent Tool Chains
Individually safe tool outputs create security risks when composed across agent sessions.
Kieran Voss-Nakamura
Machine Identity and Inventory Management for Production AI Agents
Controlling AI agents requires cataloging them before securing their access.
Desmond Okwuosa