Prompt Injection in the Wild: UAC-0099 Hid Anti-AI Text
TL;DR · the short version
Prompt injection is attacker-controlled text in a model's context that redirects what the model does. In the clearest in-the-wild case so far, ESET reported (10 September 2026) that the Russia-aligned group UAC-0099 put a comment in a malicious VBS script that reads like a request for help building a nuclear weapon, so that an AI tool analysing the file would refuse before reaching the code. The script installs MATCHBOIL, UAC-0099's own loader. Beyond that case, Google and Unit 42 have found injections on live websites aimed at AI agents, including attempts to approve scam ads, delete data and send money, but little confirmed harm. The serious results, such as data theft from Microsoft 365 Copilot and GitLab Duo and code execution through Semantic Kernel, Gemini CLI and Codex workflows, were researcher proofs of concept that the vendors fixed. Picus Security's 19 November 2026 review sums up the picture: real-world impact so far is limited, but agents connected to email, terminals or payment tools raise the stakes. The controls are the same everywhere: treat context and model output as untrusted, limit and validate tool calls, require human approval for high-impact actions, and keep testing with adaptive and obfuscated prompts. For SOC teams: a model's refusal or "clean" verdict on a file is not evidence it is safe.
On 19 November 2026, Picus Security published a review of prompt injection as it is actually being used. Its sharpest example comes from ESET: the Russia-aligned group UAC-0099 put a comment in a malicious VBS script that reads like a request for help building a nuclear weapon, so that an AI tool asked to analyse the file would refuse before it reached the code. The script installs MATCHBOIL, the group's loader. Elsewhere, the picture is many attempts, little confirmed harm, and a set of serious researcher findings that show what an attack could do to an agent with real tools.
Who this is for: CISOs, SOC and threat-intel leads who use LLMs to triage files or alerts, and anyone who runs AI agents connected to email, a terminal, a code repository or a payment tool.
Based on Picus Security's post of 19 November 2026 (Umut Bayram) and the primary sources it cites, each re-read on 20 November 2026. Only examples we could trace to their own source are included. PlayCISO is not affiliated with any company named here.
Confirmed vs not yet known
ESET reports: a UAC-0099 VBS script, used against a target in Ukraine, carried a comment written to trigger an AI model's safety guardrails; the script downloads and installs MATCHBOIL. ESET calls the technique GuardBreaker.
CERT-UA confirms separately: since mid-summer 2026 UAC-0099 has delivered VBS scripts in ZIP archives that unpack Notepad++ 8.8.3 with a malicious plugin and lead to an updated MATCHBOIL.V2. CERT-UA's advisory does not mention any AI-directed text.
Not known: whether the GuardBreaker comment ever stopped a real AI analysis tool. ESET presents it as an attempt that could work in the attackers' favour, not as a confirmed success.
Across the wider picture: Google and Unit 42 have seen injection attempts on live websites; we found no published case where one of those caused confirmed data loss or a payment. The data-theft and code-execution cases below were researcher proofs of concept, fixed by the vendors.
The UAC-0099 case, for defenders
According to ESET's WeLiveSecurity post (Tomáš Foltýn, 10 September 2026), the comment does nothing when the script runs. Its only audience is a language model. If a security team pastes the script into a chat assistant, or an automated pipeline sends it to an LLM for a verdict, the model reads what looks like a request for weapons help and may refuse the whole task. ESET classes it as a simple indirect prompt injection: the attacker's text reaches the model through the file being analysed. The script then does its real job, which is to download and install MATCHBOIL, a loader ESET says is used only by UAC-0099 to bring in further payloads.
CERT-UA first described MATCHBOIL as a C# loader used by UAC-0099 against Ukrainian government, defence and defence-industry targets, and noted it checks for running analysis tools including IDA, Wireshark, Fiddler and Procmon. ESET draws the parallel: planting text aimed at an AI model is the same idea as hiding from a debugger, applied to the tools defenders now use. In its 2026 advisory, CERT-UA documents the current chain: a phishing email with a linked image, a ZIP holding a VBS file with a double extension padded with spaces, then Notepad++ 8.8.3 side-loading a malicious NppExport.dll (LUNCHPOKE), a scheduled task running BURNYBEAR, and MATCHBOIL.V2.
Why a refusal is not a verdict
The lesson is not about one group. Any file, email or log line you hand to a model is attacker-controlled input to that model. The attacker can aim for three outcomes, and all three look harmless in a dashboard:
- Refusal. The model declines (GuardBreaker's goal). If the pipeline treats "no answer" as "nothing found", the file passes.
- Early stop. The model runs out of context or gives up part way, and only the harmless part was read. ESET mentions an npm package that repeated one phrase tens of thousands of times to fill a model's context window.
- False reassurance. Text in the file tells the model the code is safe, and the model repeats it.
What ESET recommends, and what that means in a SOC:
- Know each LLM tool's place in the decision chain: what it inspects, whether its answer can close a ticket on its own, and what happens when it refuses, errors or times out.
- Never let one model decide a file is safe. Keep static and dynamic analysis, signatures and sandbox detonation independent of the model, and cross-check with a second model or a person.
- Treat missing output as a signal. A refusal, a truncated answer or a timeout on a suspicious file should raise its priority, not lower it. Log refusals and alert on them.
- Test your own pipeline with a harmless file that carries guardrail-triggering or "this file is safe" text, and confirm it ends up in front of a human.
What has actually been seen
Picus lists about twenty examples. We kept the ones we could trace to the original publisher and split them by what they are: attempts found in the wild, and researcher demonstrations against real products.
Attempts found in the wild
| What | Who reports it | What the injection tried to do | Impact reported |
|---|---|---|---|
| UAC-0099 VBS script (GuardBreaker) | ESET, 10 Sep 2026 | Make AI-assisted analysis refuse to inspect a MATCHBOIL dropper | Not reported. ESET describes the intent, not a confirmed effect |
| Injections on public web pages | Google, 23 Apr 2026 (scan of Common Crawl) | Mostly pranks, guidance for AI summaries, SEO manipulation and notices deterring AI agents; some data exfiltration and file-deletion attempts | Google judged the exfiltration attempts rare and unsophisticated and the deletion attempts unlikely to succeed. The malicious category grew 32% (relative) between November 2025 and February 2026 |
| Web pages aimed at AI agents | Unit 42, 3 Mar 2026 | 22 payload techniques in telemetry, including getting an AI ad reviewer to approve a scam ad (24 injection attempts on one page), deleting a database, a fork bomb, steering an agent to a Stripe donation link and trying to send $5,000 to an attacker's payment account | Unit 42 says it is not aware of any confirmed case where the ad-review attempt worked against a deployed agent |
Researcher demonstrations against real products (all fixed)
| Product | Who reports it | How the text got in | What the researchers showed |
|---|---|---|---|
| Microsoft 365 Copilot | Johann Rehberger, 26 Aug 2024 | A malicious email or shared document | Copilot searched other mail on its own, pulled in data such as MFA codes, and hid it in a link using invisible Unicode Tag characters. Rehberger says the exploits stopped working after Microsoft's changes |
| GitLab Duo | Legit Security, 20 May 2025 | Hidden text in merge requests, comments, commits, issues or code | Private source code sent to an attacker's server through a rendered image tag. GitLab fixed it |
| GitHub MCP server | Invariant Labs, 26 May 2025 | A malicious issue in a public repository | The owner's agent copied private-repository data into a public pull request. Invariant calls it an architectural problem rather than a server bug |
| Microsoft 365 Copilot (EchoLeak) | NVD, CVE-2025-32711, 11 Jun 2025 | An email the victim never has to open | Information disclosure over a network ("AI command injection"). Our EchoLeak teardown walks the chain |
| Gemini CLI triage workflow | Pillar Security, 5 May 2026 | A newly opened public GitHub issue | Theft of the workflow's Git token and a path to push to the main branch. Google disabled the workflow on 17 April 2026 and fixed it in advisory GHSA-wpqr-6v78-jr5g |
| Semantic Kernel | Microsoft, 7 May 2026 | A prompt reaching an agent with the vector-store search or Python sessions plugin | Code execution on the agent host (CVE-2026-26030, fixed in Python 1.39.4) and a sandbox escape to the Windows Startup folder (CVE-2026-25592, fixed in .NET 1.71.0) |
| OpenAI Codex repository | Depi, 11 Nov 2026 | An issue title and body fed to the Codex issue-labeller workflow | Chained with a sandbox escape and cache poisoning to a point where a malicious npm release could have been published. OpenAI deployed a fix |
Read together: in the wild, attackers are experimenting and the confirmed harm is small. In the lab, the same technique against agents with real permissions reached private code, credentials, code execution and a release pipeline. The gap between the two is mostly how many agents with those permissions are in production, and that number is growing.
Useful next, if this was relevant to you
How the text gets in, and how it hides
Picus groups the delivery surfaces as: the user's own prompt; web pages and ads an agent reads; emails a copilot retrieves; uploaded documents and files; images and video frames; repository issues, comments, source files and rules files; tool and MCP responses; and, as UAC-0099 shows, comments inside malware. The hiding techniques it lists include Base64 and hex encoding, invisible Unicode (Tag characters, zero-width characters, right-to-left overrides), text hidden with CSS or zero-size fonts, instructions in other languages, payloads split across several fragments, fake system or administrator messages, and JSON-like fields that try to break out of a prompt template.
Controls, mapped to agents with email, terminal and payment tools
Picus recommends five things. Here is what each means in practice for the three kinds of agent it singles out as higher risk.
| Control | Email agent | Terminal or coding agent | Payment agent |
|---|---|---|---|
| Treat context and model output as untrusted | Every inbound email is data, never instructions. Scan retrieved mail for hidden text before it reaches the model, and do not render links or images the model produced from it without checking the domain | Issues, PR text, README and rules files, and command output are all untrusted. Read agent config files in a cloned repo before opening it with an agent | Invoices, supplier emails and web pages are untrusted. Payee details must come from your vendor master, never from what the model read |
| Limit tool permissions | Read-only by default; sending limited to internal recipients unless approved | No "YOLO" or auto-approve mode on untrusted input; scoped, short-lived tokens; no access to secrets the task does not need | Allow-listed payees and per-transaction and daily limits enforced outside the model |
| Validate tool parameters | Check recipients, attachments and forwarding rules in code against the user's own permissions | Allow-list commands and paths; block network egress except to named hosts | Validate amount, currency and payee against the original request in code |
| Require approval for high-impact actions | External sends, forwarding rules, bulk exports | Shell commands outside the workspace, pushes, publishes, workflow changes | Every new payee and every payment above a threshold |
| Keep testing, adaptively | Seed test mailboxes with encoded and multilingual injections | Run injections through issues, comments and tool output, not just the chat box | Test with injected "updated bank details" in invoices and pages |
Two of these are worth practising before you need them: deciding which agent actions need a human (the Agent Action Approval Policy Builder and the Agent Step-Up Approval Lab), and checking untrusted text before an agent reads it (Prompt Scan). A scanner is one layer, not a fix: the prompt injection explainer covers why the problem cannot be fully solved at the model and has to be contained around it.
Timeline
| Date | Event | Source |
|---|---|---|
| 26 Aug 2024 | Microsoft 365 Copilot email injection with ASCII smuggling published (researcher proof of concept) | Embrace The Red |
| 20 May 2025 | GitLab Duo remote prompt injection published; fixed by GitLab | Legit Security |
| 26 May 2025 | GitHub MCP private-repository leak published | Invariant Labs |
| 11 Jun 2025 | CVE-2025-32711 (EchoLeak, Microsoft 365 Copilot) published | NVD |
| 3 Mar 2026 | Unit 42 reports web-based injections aimed at AI agents in the wild | Unit 42 |
| 17-24 Apr 2026 | Google disables the vulnerable Gemini CLI triage workflow, then publishes GHSA-wpqr-6v78-jr5g | Pillar Security |
| 23 Apr 2026 | Google reports its Common Crawl scan: malicious injections up 32% from November 2025 to February 2026 | |
| 7 May 2026 | Microsoft publishes Semantic Kernel CVE-2026-26030 and CVE-2026-25592, both fixed | Microsoft |
| Mid-summer 2026 | CERT-UA sees UAC-0099 change tactics: VBS in ZIP, Notepad++ 8.8.3 side-loading, MATCHBOIL.V2 | CERT-UA |
| 10 Sep 2026 | ESET publishes GuardBreaker: anti-AI comment in a UAC-0099 VBS script that installs MATCHBOIL | ESET |
| 11 Nov 2026 | Depi publishes the Codex issue-labeller injection chain; OpenAI had deployed a fix | Depi |
| 19 Nov 2026 | Picus Security publishes its review of prompt injection techniques and examples | Picus Security |
What security leaders should do this week
- Find every place an LLM gives a security verdict (file triage, phishing triage, alert summaries, code review) and check what happens on a refusal, error or timeout. If any of those can close a case, change it so they escalate.
- Add a GuardBreaker-style test to your malware and phishing triage: a harmless sample carrying guardrail-triggering or "this is safe" text. It should end up with a person.
- List your agents that can send email, run commands or move money, and for each one write down which actions need a human. Turn off auto-approve modes on any agent that reads untrusted content.
- Move parameter checks out of the prompt and into code: recipients, commands, paths, payees and amounts checked against the user's own permissions.
- Review CI workflows that feed issue or PR text to an agent. The Gemini CLI and Codex cases both started from a public issue. Check what token the agent can read and what it can trigger.
- Schedule repeat testing with encoded, multilingual and split payloads, through every channel the agent reads, not just the chat box.
For the wider pattern of attackers building agents rather than targeting them, see our ARTEX agentic hacking framework coverage; for injection through a coding agent's own config files, see AI coding agent config malware.
FAQ
What did UAC-0099 do with prompt injection?
ESET reported (WeLiveSecurity, 10 September 2026) that the Russia-aligned group UAC-0099 put a comment in a malicious VBS script that reads like a request for help building a nuclear weapon. The text does nothing when the script runs. It is there so that an AI tool asked to analyse the file hits its safety guardrails and refuses, before it gets to the code. The script itself downloads and installs MATCHBOIL, a loader ESET says only UAC-0099 uses. ESET calls the technique GuardBreaker.
Did the GuardBreaker trick actually fool an AI tool?
Not as far as anyone has published. ESET describes it as an attempt that could work in the attackers' favour and does not report a confirmed case of a real AI analysis tool being derailed by it. What it does show is that this group expects defenders to put AI in the analysis path and is planning around it, the same way it already checks for analysis tools such as IDA and Wireshark.
What is MATCHBOIL?
MATCHBOIL is a C# loader that CERT-UA attributes to UAC-0099. CERT-UA describes it fetching further payloads and setting up persistence. In 2026 CERT-UA documented an updated version, MATCHBOIL.V2, delivered by a VBS script inside a ZIP archive, which unpacks a full copy of Notepad++ 8.8.3 with a malicious plugin DLL. CERT-UA's advisory does not mention the anti-AI comment; that finding is ESET's.
Is prompt injection being used in real attacks?
Yes, but mostly as attempts with little confirmed impact so far. Google scanned Common Crawl and found pranks, SEO manipulation, anti-AI notices and rare, unsophisticated exfiltration and deletion attempts, with the malicious category up 32% between November 2025 and February 2026. Unit 42 catalogued live web pages trying to make AI agents approve scam ads, delete databases or send money, and says it is not aware of any confirmed case where the ad-review attempt worked. The serious data-theft and code-execution results so far come from security researchers, and the products involved were fixed.
Why should a SOC not trust an AI model's verdict on a suspicious file?
Because the file under analysis is attacker-controlled input to the model. Text inside it can make the model refuse, stop early, or say the file is clean. A refusal, a timeout or a "benign" verdict from one model is therefore not evidence that a file is safe. ESET's advice is to know what each LLM tool inspects and what happens when it refuses or fails, never let a single model decide code is safe, cross-check with other layers and people, and treat missing output as a reason to look harder.
What are the most important controls for AI agents with email, terminal or payment access?
Treat everything in the model's context, and everything the model outputs, as untrusted. Give each agent the fewest tools and the narrowest permissions it needs. Validate every tool call's parameters in code against the user's own permissions. Require a person to approve high-impact actions such as sending email outside the organisation, running shell commands or moving money. Then keep testing with adaptive and obfuscated injections, not a fixed list, because attackers adapt.
Where does prompt injection come from?
Anywhere text reaches the model: what the user types, web pages and ads the agent reads, emails pulled in by a copilot, uploaded documents, images and video frames, repository issues, comments and source files, rules files, and the results of tool and MCP calls. It can be hidden with encoding such as Base64, invisible Unicode characters, other languages, or split into fragments that only make sense together.
Sources
- ESET WeLiveSecurity: GuardBreaker, derailing AI-assisted malware analysis with a code comment (Tomáš Foltýn, 10 September 2026)
- CERT-UA: UAC-0099, LUNCHPOKE, BURNYBEAR, updated MATCHBOIL.V2 and use of Notepad++ 8.8.3 (2026, in Ukrainian; no mention of AI-directed text)
- CERT-UA: updated UAC-0099 toolkit, MATCHBOIL, MATCHWOK, DRAGSTARE (in Ukrainian; MATCHBOIL description and analysis-tool checks)
- Google: AI threats in the wild, the current state of prompt injections on the web (Thomas Brunner, Yu-Han Liu, Moni Pande, 23 April 2026)
- Unit 42: Fooling AI agents, web-based indirect prompt injection observed in the wild (3 March 2026)
- Microsoft Security Blog: When prompts become shells, RCE vulnerabilities in AI agent frameworks (7 May 2026)
- Pillar Security: From prompt injection to supply-chain compromise on gemini-cli (5 May 2026)
- Depi: Backdooring the next release of Codex through GitHub Action pipeline exploit chaining (11 November 2026)
- NVD: CVE-2025-32711 (Microsoft 365 Copilot, published 11 June 2025)
- Invariant Labs: GitHub MCP exploited, accessing private repositories via MCP (26 May 2025)
- Legit Security: Remote prompt injection in GitLab Duo leads to source code theft (20 May 2025)
- Embrace The Red: Microsoft Copilot, from prompt injection to exfiltration of personal information (26 August 2024)
- Picus Security: What is prompt injection (OWASP LLM01), techniques, examples and defenses (Umut Bayram, 19 November 2026; the review this post follows)
As of 20 November 2026.