Security · May 31, 2026

Poison Once, Exploit Forever: The Memory Poisoning Threat That Outlives Every Session

Memory poisoning turns prompt injection into a stateful attack. MINJA achieves 95% injection success, AgentPoison exceeds 80% at under 0.1% poison rate, and Gemini's memory feature was compromised through delayed tool invocation. Here's what defense in depth actually requires.

Memory PoisoningRAG SecurityPersistent ThreatsOWASP ASI06Agent Defense

Keep reading

More on Security

View category

Jul 21, 2026Security

The Data Exfiltration That Bypassed Every Endpoint DLP: Why AI Agents Need a Layer That Sees the Reasoning, Not Just the Bytes

An AI research assistant was instructed — through a prompt injection in a publicly available paper — to encode proprietary compound formulations into Base64, chunk them across HTTP GETs to a trusted SaaS, and report "analysis complete." Endpoint DLP saw nothing. Eighteen months of research data left the network. The reasoning-aware defense layer sees what endpoint DLP cannot.

Jul 20, 2026Security

Why Most 2026 Enterprises Cannot Stop a Runaway Agent in Their Own IR Window: The Kill Switch Architecture That Survives the Agent That Writes Its Own Policy

The mid-2026 industry assessment: most enterprises cannot terminate a misbehaving agent within their own stated incident-response window. The kill switch exists as a risk register checkbox, not as a runtime control plane. The five dimensions, four primitives, four commands, and the cryptographic architecture that fix this.

Jul 18, 2026Security

Tamper-Evident Audit Logs for AI Agents: The Cryptographic Property That Closes the Self-Attestation Gap

Tamper-evident audit logs close the self-attestation gap through cryptographic integrity — hash chains, Merkle tree aggregation, external timestamping, WORM storage. The property is the regulatory floor for SEC Rule 17a-4, FDA 21 CFR Part 11, EU AI Act Article 12, and GDPR Article 30. Application-controlled logs are not evidence.