H
Howardism
Howardism · Vol. 03Plate II · No. 02

Threats, tagged.

Notes2TagThreatsOldest28 May 2026Newest28 May 2026

Every article tagged threats, newest first.

Articles tagged Threats, sorted by date, newest first.
TitleSummaryDate
Agentic Prompt InjectionDirect 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 PoisoningCorruption 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