Prompt injection has been the top entry in the OWASP Top 10 for LLM Applications since the list was created, and when OWASP published its Top 10 for Agentic Applications in December 2025, the first item, Agent Goal Hijack, cited one of the incidents below as its defining example. What changed in 2025 is that the attacks stopped being demonstrations against toy apps. They were run against the assistants that enterprises had already rolled out to every employee, and in almost every case the victim did nothing more than ask the assistant a normal question.
What the 21 incidents have in common
Read together, the incidents follow one script. An attacker places text where the assistant will later read it: an email, a calendar invite, a shared document, a public web page, a GitHub issue, a pull-request description, a web-to-lead form, a log line, even a phone notification. The text is written to look like instructions and is often hidden (white on white, inside an HTML comment, behind a spoiler tag, in a foreign language). Days later the victim asks the assistant something unrelated. Retrieval pulls the planted text into the context window, the model treats it as instruction because it has no way to tell data from instruction, and it does what the text says. What the text says is almost always the same thing: find the most sensitive data you can reach and put it in a URL, then load that URL.
The loading step is the exfiltration. A markdown image, an image proxy, a browser tool, a DNS lookup or an outbound email are all ways to make a request leave the vendor’s infrastructure carrying the data in the address. Vendors knew this and had allowlists of permitted destinations. In at least five incidents the allowlist was bypassed: Microsoft’s own Teams and SharePoint proxy domains were on Copilot’s list (EchoLeak); Azure Blob Storage was trusted by ChatGPT (AgentFlayer); Salesforce still trusted an expired domain a researcher bought for five dollars (ForcedLeak); GitHub’s Camo image proxy was used to encode data one pixel at a time (CamoLeak); and Salesforce’s URL redactor did not recognise newer top-level domains or URLs ending in brackets, so DNS carried the data out (SalesBleed).
The 2025 incidents: zero-click exfiltration from enterprise copilots
GitHub MCP server, May 2025
Invariant Labs showed that a coding agent using GitHub’s official MCP server could be hijacked by a public issue containing hidden instructions. When the owner asked the agent to look at open issues, it read the instructions, pulled data from the owner’s private repositories (the token covered both) and wrote it into a pull request on the public repository where the attacker could read it. No CVE was issued because nothing in the server was “broken”: the agent had the permissions and did as it was told. Invariant called it a toxic agent flow. [1][2]
EchoLeak, Microsoft 365 Copilot, CVE-2025-32711, June 2025
Aim Labs reported the first zero-click attack on an enterprise copilot in January 2025; Microsoft fixed it server-side in May and published an advisory on 11 June. The attacker sends an email written to read like instructions to a human, which evades Microsoft’s cross-prompt-injection classifier, and phrased so that Copilot’s retrieval will pull it in for many unrelated questions. The instruction tells Copilot to embed the most sensitive content in its context into a markdown image URL. Because SharePoint and Teams domains were on Copilot’s content-security allowlist, the browser fetched the image through a Microsoft proxy and the data left with it. CVSS 9.3, Critical. Microsoft and Aim said there was no evidence of exploitation in the wild. [3][4][5]
CurXecute, Cursor, CVE-2025-54135, August 2025
Also from Aim Labs: a message in a public Slack channel, read by Cursor’s agent through the Slack MCP server, instructed the agent to edit the project’s MCP configuration file. Cursor started newly written MCP servers before the user approved the diff, so the attacker’s command ran even when the edit was rejected. Remote code execution on the developer’s machine, CVSS 8.6, fixed in Cursor 1.3. Check Point’s related MCPoison finding is CVE-2025-54136. [6][7]
“Invitation Is All You Need”, Google Gemini, August 2025
Researchers from Tel Aviv University, the Technion and SafeBreach presented fourteen attack scenarios at Black Hat USA. Hidden instructions in a Google Calendar invite stayed dormant until the victim asked Gemini to summarise their week; some payloads only fired when the user later typed “thanks”. Gemini then called Google Home tools, opening window shutters, switching off lights and turning on a boiler in a Tel Aviv apartment, and in other scenarios sent spam, leaked location, deleted events and opened Zoom. Google deployed classifier and confirmation mitigations and said it saw no malicious use. [8][9]
AgentFlayer, ChatGPT Connectors, August 2025
Zenity Labs shared a document into a victim’s Google Drive with white-text instructions. When the victim asked ChatGPT to summarise it, the model searched the Drive for API keys and rendered a markdown image whose URL carried them. OpenAI’s client-side URL check trusted Azure Blob Storage, so the request went through and the keys were read back from Azure logs. The same research hit Copilot Studio, Salesforce Einstein and Cursor with Jira. [10][11]
Perplexity Comet, August and October 2025
Brave’s security team hid instructions behind a Reddit spoiler tag. Asked to summarise the page, Comet treated the page text as the user’s intent, navigated to the user’s Perplexity account to read their email address, opened Gmail to read a one-time code, and posted both back to Reddit as a reply. Perplexity’s first fix was found incomplete twice. In October, LayerX’s CometJacking showed that a crafted Perplexity URL could make Comet read connected Gmail and Calendar data, base64-encode it and send it to an attacker; Perplexity initially closed the report as having no security impact. Brave’s follow-up demonstrated instructions hidden in screenshots that other AI browsers also obeyed. [12][13][14][15][16]
ShadowLeak, ChatGPT Deep Research, September 2025
Radware’s attack is the one that matters most for anyone relying on network monitoring. An email with CSS-hidden instructions sat in the victim’s Gmail. When the victim asked Deep Research to analyse their inbox, the agent read the email, base64-encoded personal data and called its server-side browsing tool to an attacker URL. The exfiltration happened from OpenAI’s cloud, so nothing crossed the victim’s endpoint or corporate gateway. OpenAI fixed it in August before disclosure. Radware’s January 2026 follow-up, ZombieAgent, showed the same class of attack could plant persistent rules in ChatGPT’s memory. [17][18][19]
ForcedLeak, Salesforce Agentforce, September 2025
Noma Security submitted a lead through a public Web-to-Lead form with instructions in the description field. When an employee asked Agentforce about leads, it rendered an image pointing at a domain that was still on Salesforce’s trusted-URL list but had expired; Noma bought it for five dollars and received the CRM data. Salesforce enforced Trusted URLs for Agentforce and Einstein in September. Noma rated it CVSS 9.4. [20][21]
Gemini Trifecta, CamoLeak and Atlas, autumn 2025
Tenable’s three Gemini findings included a log-to-prompt injection: instructions placed in an HTTP User-Agent header landed in Cloud logs, which Gemini Cloud Assist summarised and obeyed. [22] Legit Security’s CamoLeak against GitHub Copilot Chat used hidden HTML comments in a pull-request description to make Copilot read private repositories and leak them letter by letter through a dictionary of GitHub image-proxy URLs; GitHub disabled image rendering in Copilot Chat in response. [23][24] NeuralTrust showed that OpenAI’s Atlas browser treated a malformed URL in its omnibox as trusted user intent, and OpenAI’s CISO called prompt injection “a frontier, unsolved security problem”. [25][26][27]
The 2026 incidents: coding agents, intake forms and the first real exploitation
Reprompt and the Gemini calendar bypass, January 2026
Varonis showed that a Copilot Personal link with a pre-filled prompt parameter, chained with a request loop that kept pulling instructions from an attacker server, gave continuous exfiltration from a single click; Microsoft patched it in the January Patch Tuesday. [28][29] Miggo Security showed a calendar-invite payload that lay dormant until the victim asked Gemini whether they were free on Tuesday, then wrote a summary of their private meetings into a new event the attacker could see. [30][31]
Clinejection, February 2026: exploited in the wild
This is the incident that moves prompt injection from demonstrated to realised. Cline, a popular open-source coding agent, had added a GitHub issue-triage workflow that put the issue title verbatim into the prompt sent to the model. A crafted title hijacked the workflow. From there the attacker poisoned the GitHub Actions cache, reached the release pipeline and obtained the npm publish token. On 17 February 2026 a malicious version of the Cline package was published and, for roughly eight hours, every machine that updated ran a post-install script that installed an unwanted agent. Cline moved to OIDC trusted publishing afterwards. Snyk and the Cloud Security Alliance both describe it as the first documented case of indirect prompt injection against an AI agent being exploited to compromise a software supply chain. [32][33][34]
RoguePilot, GrafanaGhost, ShareLeak and PipeLeak, February to April 2026
Orca Security’s RoguePilot used a hidden comment in a GitHub issue so that, when a developer opened a Codespace from it, Copilot’s coding agent wrote the Codespace token into a commit. [35][36] Noma’s GrafanaGhost planted instructions in log lines that Grafana’s AI assistant later summarised; a weak URL check let a rendered image carry internal data out. [37][38] Capsule Security’s ShareLeak (CVE-2026-21520, Copilot Studio) and PipeLeak (Agentforce) both fed instructions through public business forms, the same path ForcedLeak had used six months earlier. [39][40]
“Comment and Control” and the CI agents, April to August 2026
Johns Hopkins researchers showed that a pull-request title, issue or comment crafted to break out of the agent’s prompt template could make Claude Code Security Review, the Gemini CLI GitHub Action and the GitHub Copilot agent run shell commands in CI, read the API keys and tokens in the environment and send them out through GitHub’s own API, so the platform became the command channel. The Claude Code variant was rated CVSS 9.4. Microsoft Threat Intelligence separately showed Claude Code Action’s file-read tool could read the process environment; Anthropic fixed it in Claude Code 2.1.128 in May 2026. All three vendors patched. [41][42][43]
Voice and CRM, June to September 2026
SafeBreach’s Fake Context Alignment attack put malicious text in WhatsApp, Slack and SMS notifications; when an Android user asked Gemini to read notifications hands-free, the text entered the context and could impersonate contacts in spoken replies, trigger smart-home actions and poison long-term memory. [44][45] Zenity’s SalesBleed, disclosed in September 2026 after Salesforce fixed it in August, again started from a Web-to-Lead form and this time encoded CRM data into DNS hostnames, defeating the Trusted-URL redactor that ForcedLeak had prompted Salesforce to add. [46][47]
By the numbers
| Figure | Source | Date |
|---|---|---|
| 21 publicly disclosed prompt-injection incidents against production assistants; 3 with CVEs (CVE-2025-32711, CVE-2025-54135, CVE-2026-21520) | This report | Nov 2026 |
| 32% of organisations had an attack on AI applications that exploited the prompt in the prior 12 months (survey of 302 security leaders) | Gartner | Sep 2025 |
| 540% year-on-year rise in valid prompt-injection reports on bug-bounty programmes | HackerOne | Oct 2025 |
| 32% increase in malicious indirect-prompt-injection content on the public web, Nov 2025 to Feb 2026, with “no particularly sophisticated attacks” seen | Google Online Security Blog | Apr 2026 |
| Indirect injections needed significantly fewer attempts to succeed than direct attacks | Lakera, Q4 2025 Agent Security Trends | Jan 2026 |
| Claude for Chrome: attack success cut from 23.6% to 11.2% with mitigations across 29 scenarios (vendor self-evaluation) | Anthropic | Aug 2025 |
| 8 major GenAI exploits in Q1 2026, 1 with a CVE; LLM01 the most-cited category | OWASP GenAI Exploit Round-up | Apr 2026 |
Two things in that table deserve emphasis. First, the CVE count. Most vendors fixed server-side and did not assign a CVE, which means these incidents do not appear in the vulnerability feeds most security teams monitor. A team tracking CVEs alone would have seen three of twenty-one. Second, the prevalence figures are from surveys, bug-bounty submissions and web scans, not from confirmed breach data. Nobody yet has a reliable count of how often this is happening unobserved. [48][49][50][51][52][53]
What has actually worked, and what has not
- Prompt-level classifiers. Every major vendor runs one. EchoLeak evaded Microsoft’s by writing the payload as instructions to a human. Classifiers raise the cost of an attack; they do not end it. Perplexity’s own write-up concedes that prompt injection “remains an unsolved problem across the industry”. [15]
- Application egress allowlists. Bypassed in at least five incidents, each time because something on the list was reachable by the attacker. An allowlist maintained by the same team that built the assistant inherits that team’s blind spots.
- Disabling rendering. GitHub’s fix for CamoLeak was to stop Copilot Chat rendering images at all. It is the only fix in this set that removes the exfiltration leg outright rather than filtering it, and it is a product cost not every vendor will accept.
- Confirmation before consequential actions. Google, Perplexity and Anthropic all added user confirmation for high-impact tool calls. It works until users click through by habit, and it does nothing for read-and-exfiltrate attacks that need no tool call.
- Architectural separation. Google DeepMind’s CaMeL design (March 2025) splits a privileged model that plans from the user’s request and a quarantined model that parses untrusted content into typed values, with a policy interpreter between them. It solved 77% of AgentDojo tasks with provable security against 84% undefended. Microsoft’s Spotlighting (2024) marks untrusted text so the model can tell it apart and cut attack success from over 50% to under 2% on GPT-family models. Neither is widely deployed in the products above. [54][55][56]
- Least privilege and single-purpose tools. The GitHub MCP, RoguePilot and Comment-and-Control attacks all succeeded because one token covered more than the task needed. OWASP’s ASI01 guidance says the same.