Early-Signal Detector

phenomenon · first seen 4 Oct 2026

prompt injection

Also: IPI, context poisoning, jailbreak prompting, instruction injection, prompt injection defense, hidden instruction attack, indirect prompt injection

Attacks where malicious instructions embedded in input data cause a language model agent to deviate from its intended task.

Prompt injection covers attacks where instructions hidden in data an LLM reads, such as web pages, code comments, tool outputs or agent metadata, override the task the model was given. The sources mostly extend it to new surfaces rather than introduce it. Academic papers cover staged injection against Claude Code and Codex, provider-side injection in proactive agents, Agent Card manipulation in the A2A protocol, compliance artifacts, and deep research agents. Developer discussion points to hidden LLM instructions on WhatsApp help pages and a Brazilian court sanctioning lawyers for injected text in a filing. The GitHub activity is mostly red-team tooling, including mirrored DeepSeek jailbreak plugins and one crypto-bot repo with keyword stuffing.

The idea has been widely discussed since 2022, so this rise reflects steady research volume as agents gain tool access, not a new concept. The live work is in agent-specific attack surfaces and defenses at trust boundaries.

Analyst note (established), 17 Nov 2026.

Mentions per week

03719 May11 Aug10 NovPapersCodeDiscussion

PapersCodeDiscussion

Rank 27 on 14 Nov 2026 · 126 mentions in 28 days · 126 independent groups · papers → code → discussion · first seen 6 weeks ago · acceleration +0.03

Earliest sources

Latest sources

Mentioned together with

model context protocolretrieval-augmented generationmulti-agent LLM systemred teamingconsole jailbreakinglinear probingtool callingjailbreak attackAgentDojoadversarial training