Howardism · Vol. 03Plate II · No. 02
Threats, tagged.
Notes2TagThreatsOldest28 May 2026Newest28 May 2026
Every article tagged threats, newest first.
| Title | Summary | Date |
|---|---|---|
| Agentic Prompt Injection | Direct and indirect injection of malicious instructions into an agent; LLMs cannot reliably distinguish information from instructions; defenses are spotlighting (50%→<2%), constitutional classifiers (95% blocked), input isolation, and attack-surface reduction | |
| Memory and Context Poisoning | Corruption of persistent agent memory that influences behavior long after the initial injection; includes RAG poisoning, shared-context poisoning, and slow long-term memory drift; defended via memory isolation, integrity validation, and retention policies |