Field report N-43 / J1 · sheet /operations

Operations & RestructuringFiled from the territory

Hidden Text Can Hijack Corporate AI Assistants — and No Complete Fix Exists

Prompt injection lets hidden text in emails and web pages hijack corporate AI assistants with no code or stolen passwords. There is no complete fix; containment is the only real defense today.

Grid
N-43 / J1
Surveyed
November 20, 2026
Reading
4 min read
Correspondent
Sophie Lindqvist
How hidden text in emails and web pages can hijack a company AI assistant
Fig. N-43 / J1How hidden text in emails and web pages can hijack a company AI assistant — Nile / pixabay

Field notes

  1. Simon Willison coined 'prompt injection' in September 2022; he named the 'lethal trifecta' risk in June 2025.
  2. CVE-2025-32711 (EchoLeak) was a zero-click flaw in Microsoft 365 Copilot, reported by Aim Security and fixed server-side.
  3. In February 2023, Stanford student Kevin Liu extracted Bing Chat's system prompt, including the codename 'Sydney', via prompt injection.
  4. Invariant Labs demonstrated MCP tool-poisoning in April 2025 and a GitHub MCP server attack in May 2025.
  5. No complete fix exists; Google DeepMind's 2025 CaMeL design remains a research proposal, not a deployed solution.

By Sophie Lindqvist

Prompt injection — an attack in which hidden instructions in emails, web pages or files hijack an AI assistant — has no complete fix today. Every company running a language-model tool faces the same structural weakness: the model cannot reliably tell its developer's commands apart from text a stranger wrote.

Developer Simon Willison coined the term in September 2022, modeling it on SQL injection, where an attacker types database commands into an ordinary input field. The parallel is direct. But SQL injection got a structural remedy — parameterized queries (a technique that sends command and user input to the database separately). Language models have no equivalent, because instructions and data arrive in the same human language.

What does this change for corporate security teams? It shifts the question from "can our assistant be tricked?" to "what can a tricked assistant do?" The working assumption becomes: it will be tricked, plan for that moment.

How the attacks work

Direct injection is typed into the chat window by the user; indirect injection hides inside content the AI reads on the user's behalf. One of the earliest public examples came in September 2022, when users took over a Twitter bot run by remote-work job site remoteli.io — built on OpenAI's GPT-3 — using plain English. No code, no stolen password, no technical skill.

Kai Greshake and co-authors formalized indirect injection in a February 2023 paper, "Not what you've signed up for." For companies, this is the larger concern: the attacker never touches the victim's chat. Every source an assistant pulls from — inbound email, shared documents, search results, third-party websites — becomes an entry point.

Why can't models separate instructions from data? Everything arrives as a single stream of tokens (small chunks of text the model processes) in the context window. Position and formatting offer hints learned in training, not enforced rules. OpenAI's 2024 "instruction hierarchy" method teaches models to favor developer instructions over tool outputs; it reduces the problem without removing it.

Data theft, zero-click

Prompt injection can turn a chat window into an exfiltration channel. A planted instruction tells the model to append private data to an image URL's query string (the part of an address after a question mark). The chat interface, rendering markdown (a lightweight formatting language), fetches the image automatically — no click needed. Security researcher Johann Rehberger demonstrated variants across several AI products.

The 2025 EchoLeak flaw, tracked as CVE-2025-32711, brought this pattern into Microsoft 365 Copilot. Reported by Aim Security and described as zero-click, it required no opening or clicking by the victim. Microsoft fixed it server-side. That removes the safeguard companies usually lean on: employee vigilance.

System prompts (the hidden instructions a company gives its model) are also extractable. In February 2023, Stanford student Kevin Liu used prompt injection to pull Bing Chat's initial instructions, including the codename "Sydney" — through the ordinary chat interface, with text as the only weapon. The design rule that follows: assume the system prompt will be published, and keep credentials, API keys and pricing logic out of it.

The lethal trifecta

Willison named the "lethal trifecta" in June 2025: an agent with access to private data, exposure to untrusted content, and the ability to communicate externally. One injected instruction can then read sensitive files and send them to an attacker. Remove any one of the three, and this theft route closes. Tool access raises the stakes further — an injected instruction runs with the agent's permissions; if it can send email, so can the attacker.

The Model Context Protocol (MCP), an open standard Anthropic introduced in November 2024, adds a supply-chain dimension. Invariant Labs showed in April 2025 that malicious instructions hidden in an MCP tool's description can steer agents; in May 2025 it reported a related attack using the official GitHub MCP server, where a malicious repository issue could lead an agent to leak private-repository data. Each MCP server deserves review like any software dependency.

Injection is not jailbreaking

Willison draws the distinction: jailbreaking overrides a model's safety training; prompt injection overrides an application's logic. A model that refuses every harmful request can still follow an injected instruction to forward an email — breaking no safety rule, only the application's intent. The test is short: a jailbreak changes what a model will say; an injection changes whom it works for.

What companies can actually do

  • Least-privilege access: a meeting summarizer does not need to send email; a coding assistant does not need payroll records.
  • Human approval before irreversible actions — sending email, moving money, deleting files.
  • Limited outbound channels to approved addresses and domains.
  • Logging tool calls to reconstruct what a hijacked assistant read and sent.
  • Repeated red-teaming after every model or tool change; new versions respond differently to the same attack.
  • Input and output filters, which catch known patterns but never all reworded variants.

Google DeepMind's 2025 CaMeL design separates trusted instructions from untrusted data at the system level, without relying on the model to tell them apart. It is a research proposal, not a deployed fix. Until something like it survives production, containment — not a cure — is what companies actually have.

via qz.com (Source)

Share this article:

More from Sophie Lindqvist

Sophie Lindqvist

Show full bio

Correspondent covering money, funding and market shifts around corporate strategy and chief-executive leadership at CEO Frontier.

515 articles

Filed under

Adjacent sheets

« Previous article

End of sheetN-43 / J1 · November 20, 2026