---
title: AI that reads requests must not obey them | Disruptive Lab
description: When AI reads e-mails, tickets or forms and acts on them, a stranger's text can give it orders. Four design rules that keep the process in your hands.
url: https://disruptivelab.dev/en/news/ai-reads-requests
language: en
covers: 2026-03-02/2026-03-08
---

1.  [News](https://disruptivelab.dev/en/news)
2.  [Data, integration, automation](https://disruptivelab.dev/en/news/data-and-integration)
3.  AI that reads requests must not obey them

[Guide](https://disruptivelab.dev/en/news/guides)  [Data, integration, automation](https://disruptivelab.dev/en/news/data-and-integration)

# AI that reads requests must not obey them

An AI step that reads requests from outside will sometimes take a stranger's text for an order, so the process must be designed so that it cannot act on that order alone.

- **Published:** 17 November 2026
- **Updated:** 17 November 2026
- **Reading time:** 8 min read
- **Editor:** David Lahuta

Week of 2 March 2026

## In short

-   From December 2025 the open source project Cline let an AI model sort its incoming bug reports and gave it the right to run commands.[1](https://disruptivelab.dev/en/news/ai-reads-requests#source-1)[3](https://disruptivelab.dev/en/news/ai-reads-requests#source-3)
-   In a researcher's tests the title of a public report was enough to make that AI step install code chosen by an outsider, and on 17 February 2026 a stolen publishing key put out a changed version of the project's tool, downloaded about 4,000 times.[1](https://disruptivelab.dev/en/news/ai-reads-requests#source-1)[3](https://disruptivelab.dev/en/news/ai-reads-requests#source-3)[4](https://disruptivelab.dev/en/news/ai-reads-requests#source-4)
-   The security community's list of AI risks says no reliable way exists today to stop text from giving a model orders.[6](https://disruptivelab.dev/en/news/ai-reads-requests#source-6)
-   The answer is in the design of the process. The AI reads and proposes, a person approves what has consequences, no secret is within its reach and every step is recorded.

## Contents

1.  [What happened, as the researchers and the project report it?](https://disruptivelab.dev/en/news/ai-reads-requests#what-happened-as-the-researchers-and-the-project-report-it)
2.  [Why can text from outside give an AI step orders?](https://disruptivelab.dev/en/news/ai-reads-requests#why-can-text-from-outside-give-an-ai-step-orders)
3.  [Which design rules keep the AI step safe?](https://disruptivelab.dev/en/news/ai-reads-requests#which-design-rules-keep-the-ai-step-safe)
4.  [What should you check in each automated process?](https://disruptivelab.dev/en/news/ai-reads-requests#what-should-you-check-in-each-automated-process)

## What happened, as the researchers and the project report it?

Cline is an open source assistant for software developers, distributed as an editor extension and as a command line tool through npm, the public registry for such packages.[1](https://disruptivelab.dev/en/news/ai-reads-requests#source-1)[3](https://disruptivelab.dev/en/news/ai-reads-requests#source-3) On 21 December 2025 its maintainers added an automated workflow on GitHub that sorted and answered new issues, the public bug reports and requests anyone can open on the project's page.[1](https://disruptivelab.dev/en/news/ai-reads-requests#source-1)[3](https://disruptivelab.dev/en/news/ai-reads-requests#source-3) The workflow handed each new issue to an AI model from Anthropic.[1](https://disruptivelab.dev/en/news/ai-reads-requests#source-1)[3](https://disruptivelab.dev/en/news/ai-reads-requests#source-3) Two settings made it dangerous. Any GitHub user could start it by opening an issue, and the AI was allowed to run commands.[1](https://disruptivelab.dev/en/news/ai-reads-requests#source-1)

The researcher Adnan Khan found that the title of an issue was copied straight into the AI's instructions.[3](https://disruptivelab.dev/en/news/ai-reads-requests#source-3) In his tests, a title that claimed a tool was missing and asked the AI to install a "helper tool" first was enough.[3](https://disruptivelab.dev/en/news/ai-reads-requests#source-3) The AI ran the install, and with it a script the attacker controlled.[3](https://disruptivelab.dev/en/news/ai-reads-requests#source-3) From there an attacker could plant files in storage shared with the project's nightly release workflow, which held the keys for publishing new versions.[1](https://disruptivelab.dev/en/news/ai-reads-requests#source-1)[3](https://disruptivelab.dev/en/news/ai-reads-requests#source-3)

Date

What happened

21 December 2025

The AI workflow for sorting issues is added.[1](https://disruptivelab.dev/en/news/ai-reads-requests#source-1)

1 January 2026

The researcher reports the weakness privately through GitHub. It is not addressed until he publishes it.[1](https://disruptivelab.dev/en/news/ai-reads-requests#source-1)

9 February 2026

He publishes it. Within 30 minutes the project removes the workflow and starts replacing its keys.[1](https://disruptivelab.dev/en/news/ai-reads-requests#source-1)

9 February 2026

The wrong npm key is deleted and the exposed one stays valid.[1](https://disruptivelab.dev/en/news/ai-reads-requests#source-1)[3](https://disruptivelab.dev/en/news/ai-reads-requests#source-3)

17 February 2026, 3:26 Pacific time

An unauthorised party uses that key to publish version 2.3.0 of the command line tool, with one added line that installs a separate AI agent, OpenClaw.[1](https://disruptivelab.dev/en/news/ai-reads-requests#source-1)[4](https://disruptivelab.dev/en/news/ai-reads-requests#source-4)

17 February 2026, 11:40 UTC

A security firm's monitoring flags the release, about 14 minutes after it went out.[4](https://disruptivelab.dev/en/news/ai-reads-requests#source-4)

17 February 2026, 11:30 Pacific time

Version 2.3.0 is withdrawn, after about eight hours.[2](https://disruptivelab.dev/en/news/ai-reads-requests#source-2)

The changed version was downloaded about 4,000 times.[4](https://disruptivelab.dev/en/news/ai-reads-requests#source-4) The researcher writes that the party behind it was not him but somebody who found his proof on his test copy and used it against the real project.[3](https://disruptivelab.dev/en/news/ai-reads-requests#source-3) The project states that the installed agent is a legitimate open source program, that no malicious code was delivered and that no user data was accessed.[1](https://disruptivelab.dev/en/news/ai-reads-requests#source-1) It also draws the lesson itself. Giving a language model the right to run commands where it processes text from anyone is, in its words, functionally the same as giving every GitHub user that right.[1](https://disruptivelab.dev/en/news/ai-reads-requests#source-1) On 5 March 2026 the security company grith summed the case up as five steps from a GitHub issue title, with natural language as the entry point.[5](https://disruptivelab.dev/en/news/ai-reads-requests#source-5)

## Why can text from outside give an AI step orders?

Because a language model cannot reliably tell instructions from data. The OWASP Top 10 for LLM Applications 2026, the security community's list of the main risks of systems built on AI language models, ranks this first and calls it prompt injection, any input that changes what the model does in a way its builder did not intend.[6](https://disruptivelab.dev/en/news/ai-reads-requests#source-6) It explains that the model receives its instructions and the content it reads as one stream, with no built in boundary between them.[6](https://disruptivelab.dev/en/news/ai-reads-requests#source-6)

The list names the places such text comes from, and they are ordinary business inputs.[6](https://disruptivelab.dev/en/news/ai-reads-requests#source-6) Public web pages and e-mails from unknown senders are untrusted.[6](https://disruptivelab.dev/en/news/ai-reads-requests#source-6) Issue titles in a public bug tracker are only partly trusted.[6](https://disruptivelab.dev/en/news/ai-reads-requests#source-6) Even a company's own documents and mail can carry a planted text, for example one that came in through a public form.[6](https://disruptivelab.dev/en/news/ai-reads-requests#source-6) The attacker does not need to break into anything. He puts text where the AI will read it, and the AI does the work with the company's own rights.[6](https://disruptivelab.dev/en/news/ai-reads-requests#source-6)

The list is plain about the cure. No reliable way to prevent prompt injection exists today, so the system around the model should be designed on the assumption that its instructions will one day be bypassed, and should limit what the model may do.[6](https://disruptivelab.dev/en/news/ai-reads-requests#source-6)

For a company this means one question for every process where AI reads something from outside, an incoming e-mail, a support ticket, a web form, an attachment from a supplier. If someone wrote an order into that text, what could the AI step do with it?

## Which design rules keep the AI step safe?

Four rules cover most of it. Each one rests on the OWASP guidance and on what the project changed after the incident.[6](https://disruptivelab.dev/en/news/ai-reads-requests#source-6)[1](https://disruptivelab.dev/en/news/ai-reads-requests#source-1)

**The AI reads and proposes, it does not act on its own.** OWASP advises giving an AI step only the tools its task needs.[6](https://disruptivelab.dev/en/news/ai-reads-requests#source-6) Its own example is a step that summarises e-mails and needs to read them, but not to send or delete them.[6](https://disruptivelab.dev/en/news/ai-reads-requests#source-6) It also advises against open ended tools such as running any command, in favour of narrow ones that do one thing.[6](https://disruptivelab.dev/en/news/ai-reads-requests#source-6) Cline's workflow could run commands, and the project now says any future automation will only read, with no right to run commands.[1](https://disruptivelab.dev/en/news/ai-reads-requests#source-1)

**A person approves every action with consequences.** OWASP recommends that a person approves high impact actions before they happen.[6](https://disruptivelab.dev/en/news/ai-reads-requests#source-6) The person should see the exact action that will run, not a summary of it, and OWASP warns that too many approvals dull the judgement of whoever gives them.[6](https://disruptivelab.dev/en/news/ai-reads-requests#source-6) So the split matters. Steps that are easy to undo may run alone, while irreversible ones, such as a payment, go to a person.[6](https://disruptivelab.dev/en/news/ai-reads-requests#source-6) Where one AI step reads untrusted input, reaches sensitive data and can change something or send something out, OWASP asks for a person's approval of every action.[6](https://disruptivelab.dev/en/news/ai-reads-requests#source-6)

**No secret is within its reach.** OWASP advises keeping keys and the power to change things in ordinary application code, outside the model, and letting each action run with the rights of the person it is done for, no more.[6](https://disruptivelab.dev/en/news/ai-reads-requests#source-6) The case shows why "within reach" means more than what the AI is handed. The sorting workflow itself held no publishing key, but it shared storage with the release workflow that did.[1](https://disruptivelab.dev/en/news/ai-reads-requests#source-1) The project has since stopped using that shared storage in every workflow that handles publishing keys.[1](https://disruptivelab.dev/en/news/ai-reads-requests#source-1)

**Every step is recorded, and someone reads the record.** OWASP advises logging and watching what an AI step does, so that unwanted actions are found and answered.[6](https://disruptivelab.dev/en/news/ai-reads-requests#source-6) Records only help when somebody looks at them. The researcher found failed runs of the nightly release between 31 January and 3 February 2026 that bore the trace of his own tool, and writes that the failures were tolerated rather than investigated.[3](https://disruptivelab.dev/en/news/ai-reads-requests#source-3) The project now also checks a replaced key against the key itself, not only against a management screen.[1](https://disruptivelab.dev/en/news/ai-reads-requests#source-1)

Rule

In a business process

What the project changed afterwards

The AI reads and proposes

It sorts e-mails and drafts answers but sends nothing itself

Future automation only reads, with no right to run commands[1](https://disruptivelab.dev/en/news/ai-reads-requests#source-1)

A person approves

A change to data, a payment or a message out waits for a named person

The AI workflows were removed entirely[1](https://disruptivelab.dev/en/news/ai-reads-requests#source-1)

No secret within reach

The AI step has no passwords or keys and no shared storage with steps that do

No shared storage in workflows that hold publishing keys[1](https://disruptivelab.dev/en/news/ai-reads-requests#source-1)

Every step recorded

Each input, proposal, approval and action is logged and failures are looked into

A replaced key is checked against the key itself[1](https://disruptivelab.dev/en/news/ai-reads-requests#source-1)

This guide stays on the design of the process. Which data an AI may see and where it runs are separate questions with their own answers.

## What should you check in each automated process?

Go through every process in which AI reads something that came from outside, once before it goes live and again after each larger change.

Before you decide

-   Which inputs does the AI step read, and can a stranger write any of them, an e-mail, a form, a ticket or an attachment?
-   What can it do besides reading and proposing? Every tool and every right is listed.
-   Which of its actions change data, send something out or move money, and who approves each one before it happens?
-   Does the person who approves see the exact action, not a summary?
-   Which passwords, keys and data can it reach, directly or through storage it shares with other steps?
-   Does it act with the rights of the person it works for, and no more?
-   Is every input, proposal, approval and action recorded, and does someone look into unexplained failures?
-   Has it been tested with hostile inputs by someone who knows how its defences work, as OWASP advises?[6](https://disruptivelab.dev/en/news/ai-reads-requests#source-6)
-   Who receives a report of a weakness, and how soon does someone answer it?
-   Can the AI step be switched off quickly without stopping the whole process?

## What this means for a company in Europe

E-mails, tickets and forms usually carry personal data, so the General Data Protection Regulation applies to the process that reads them. Its Article 25 asks the controller to build appropriate technical and organisational measures into the processing, both when the means are chosen and while it runs.[7](https://disruptivelab.dev/en/news/ai-reads-requests#source-7) By default, only the personal data each purpose needs may be processed.[7](https://disruptivelab.dev/en/news/ai-reads-requests#source-7) Article 32 asks for security appropriate to the risk, including the ability to keep systems confidential and intact, and a process for regularly testing whether the measures work.[7](https://disruptivelab.dev/en/news/ai-reads-requests#source-7) It names unauthorised disclosure of and access to personal data among the risks to weigh.[7](https://disruptivelab.dev/en/news/ai-reads-requests#source-7) In our view the four rules above are the kind of measures these articles describe. Whether they are enough for a given process is for each controller to assess.

## Next step

If it would help, we are glad to look with you in a short meeting at one process in your company where AI reads incoming requests, and at what it may do on its own. The AI Business Harness is built on the same rules, with each user's permissions, a person approving changes and a record of every step. What you do with it afterwards is entirely your decision.

[AI Business Harness](https://disruptivelab.dev/en/ai-business-harness)

## Sources

1.  1. [Cline, Post-mortem, unauthorized Cline CLI npm publish on February 17, 2026 (24 February 2026)](https://cline.bot/blog/post-mortem-unauthorized-cline-cli-npm). Read on 17 November 2026
2.  2. [Cline, security advisory GHSA-9ppg-jx86-fqw7, unauthorized npm publish of cline@2.3.0 (17 February 2026)](https://github.com/cline/cline/security/advisories/GHSA-9ppg-jx86-fqw7). Read on 17 November 2026
3.  3. [Adnan Khan, Clinejection, compromising Cline's production releases just by prompting an issue triager (9 February 2026, with later updates)](https://adnanthekhan.com/posts/clinejection/). Read on 17 November 2026
4.  4. [StepSecurity, Cline supply chain attack detected, cline@2.3.0 silently installs OpenClaw (17 February 2026)](https://www.stepsecurity.io/blog/cline-supply-chain-attack-detected-cline-2-3-0-silently-installs-openclaw). Read on 17 November 2026
5.  5. [grith, A GitHub issue title compromised 4,000 developer machines (5 March 2026)](https://grith.ai/blog/clinejection-when-your-ai-tool-installs-another). Read on 17 November 2026
6.  6. [OWASP GenAI Security Project, Top 10 for LLM Applications 2026, LLM01 Prompt Injection and LLM03 Excessive Agency](https://genai.owasp.org/resource/owasp-genai-llm-top-10-2026/). Read on 17 November 2026
7.  7. [Regulation (EU) 2016/679, General Data Protection Regulation, Articles 25 and 32, Official Journal of the EU](https://eur-lex.europa.eu/eli/reg/2016/679/oj/eng). Read on 17 November 2026

## Change log

1.  17 November 2026 Published.

## Related articles

Briefing · Data, integration, automation

### [AI shopping agents at your e-shop's door](https://disruptivelab.dev/en/news/ai-shopping-agents)

eBay bans AI agents that buy for people, Google offers an open standard to connect them. What blocking or connecting means for your order system.

Week of 19 January 2026

Guide · European rules

### [The AI Act for companies that only use AI](https://disruptivelab.dev/en/news/ai-act-for-deployers)

A company that uses AI but builds none is a deployer under the AI Act. What applies already, what comes in December 2027 and what to keep on record.

17 November 2026

## Company details

- **Company name:** Disruptive Lab s.r.o.
- **Registered office:** Vavrečkova 5262, 760 01 Zlín
- **Company ID (IČO):** 08240647
- **VAT ID (DIČ):** CZ08240647
- **E-mail:** office@disruptivelab.dev
- **Phone:** +420 777 702 032
