Signal SentrySearchUpdated Nov 20, 05:30 UTCAM Drop
Security lens · AI agents

Treat outside text as untrusted input

Instructions hidden in a web page, file or message get less power over an AI agent. No method blocks them all, so this limits the damage.

Where to start

Our reading of the guidance below, for a small team. A starting point, not a complete program.

  1. Keep fetched pages, files and messages apart from the instructions an agent follows, and mark them as untrusted.
  2. Give an agent that reads outside content no tools that can send, delete, pay or publish.
  3. Before release, test agents against content that carries planted instructions.

The guidance

  1. OWASP · LLM01:2025 Prompt Injection

    Separate and clearly denote untrusted content to limit its influence on user prompts.
  2. OWASP · LLM Prompt Injection Prevention Cheat Sheet

    Keep trusted instructions separate from untrusted data, but do not treat text labels or prompt wording as an enforcement boundary.

Short quotes from the official guidance, linked. Read the source before you act on it.

Incidents from the Drops

Security

Preprint: A2A Agent Cards can be rigged to sway which agent an orchestrator picks

Researchers probed 36.1 million domains, found 310 deployed Agent Cards and built SECCARD. Its best LLM judge caught 97.0% of injected attacks at a 1.7% false-positive rate; on another test nine judges ranged 25.8% to 92.3%.

Why it matters: Text in an agent's public description can steer selection before any call is made. Treat Agent Card fields as untrusted input, and vet the judge model you rely on.

Preliminary Way in: attacker-written text in an A2A Agent Card that sways which agent gets chosen

Controls that would have helped:

On the Frontier watch →From Drop #092 · Tue, Nov 17, 2026 · Evening
Security

CISA: prompt injection in Jet Admin Cloud Platform can leak .env credentials

CISA published advisory VA-26-317-01: a prompt injection flaw (CVE-2026-123965, CVSS 9.9) in all versions of Jet Admin Cloud Platform lets anyone holding a free account remotely expose .env files and steal credentials.

Why it matters: An AI feature that can read server config turns one prompt into a credential leak. If you run internal-tool builders, check which files their AI can reach.

Confirmed Way in: a prompt injection from a free account that exposed .env configuration files

Controls that would have helped:

From Drop #085 · Sat, Nov 14, 2026 · Morning
Security

Preprint: every agent memory system tested could be poisoned across sessions

Researchers report in a preprint that MemoryDojo, 155 multi-session trajectories across five memory systems, found all of them susceptible to instructions planted in one session and acted on later, with wide variation in effects.

Why it matters: If your agents remember across sessions, treat stored memories as untrusted input and test what a planted note can make them do days later.

Preliminary Way in: malicious instructions saved to agent memory in one session, then retrieved and acted on later

Controls that would have helped:

On the Frontier watch →From Drop #076 · Mon, Nov 9, 2026 · Evening