Prompt injection
Hiding instructions in content a model reads (web pages, emails, documents, tool output) so it follows the attacker instead of the user.
Prompt injection smuggles instructions into data a model processes: a web page, an email, a retrieved document or a tool result. Because LLMs don’t reliably separate instructions from data, the model may follow the attacker. For agents with tools and credentials, injection becomes an action: in the Cornell Tech/Yale study, agents fetched emailed approval codes and raised their own spend limits, and AgentCorruption started with one prompt to a public agent. Hidden WhatsApp help-page prompts show the technique used by a platform on its own content. India’s AI auditing guidelines make it a named audit area.
Timeline
Every dated event in the wiki that involves Prompt injection, newest first.
- Google releases Gemini 4 Hydrogen
A cheaper Gemini 4 tier at $1.50 / $7.50 per million tokens with a 1M-token context. Google cites 51 on the Artificial Analysis Intelligence Index (Argon: 50) and 2.3x more resilience on Gray Swan's indirect Prompt injection benchmark (vendor claims).
- IBM previews Granite LLMON models against Prompt injection
Experimental 3B and 8B Granite 4.1 LLMON previews built around an LLM-native markup language; IBM reports near-zero attack success on the OPI benchmark with general capability on par with Granite 4.1 Instruct (vendor claim; not for production).
Sources: IBM Granite (Hugging Face blog), [link]
- India CERT-In → AI systems and their auditors (India)
Published 107-page AI Security Auditing Guidelines: four risk tiers (low, moderate, high, critical), independent third-party audits for high-risk systems, more extensive assessment and continuous monitoring for critical ones, and coverage of threats such as prompt injection, data poisoning and excessive AI permissions; builds on the existing six-hour incident-reporting rule and 180-day log retention. Not a new AI law.
Full article: India AI Security Auditing GuidelinesSources: The420.in, “CERT-In Issues AI Security Audit Guidelines, Sets Four Risk Categories”
- Microsoft patches a Copilot flaw that let it rewrite its own MCP config (Visual Studio Copilot CVE-2026-103572)
A trailing dot in a file name slipped past Copilot's path check in Visual Studio and SQL Server Management Studio, letting the agent edit its own MCP and settings files and run commands, including via Prompt injection. Microsoft rated CVE-2026-103572 an Important RCE and fixed it on Patch Tuesday but paid no bounty, the researcher says.
Sources: Embrace The Red, “Visual Studio Copilot and SSMS Remote Code Execution (CVE-2026-103572)”
- Prompt-injection flaw in Opera Neon reported by Brave
Hidden HTML on a page could steer Neon's AI into misusing browser tools, an indirect prompt injection. Opera says it fixed the issue and that the risk cannot be reduced to zero.
Sources: Brave, “Prompt injection flaw in Opera Neon”; Opera, [link]
Connections
Prompt injection and everything one link away. Drag to move, scroll to zoom, click a node to open it.Open in the full graph →
What links here
Incident Codex GitHub Actions prompt-injection chain · Visual Studio Copilot CVE-2026-103572
Concept Context window · Retrieval-augmented generation (RAG)
Standard India AI Security Auditing Guidelines
Protocol LlamaIndex Document Parsing Protocol
Product Opera Neon · Strawberry (browser agent)
Sources
- Wikipedia, “Prompt injection”