AI bug bounty case files: real findings, taken apart
Five publicly disclosed vulnerabilities in real AI products. For each one I show the chain in three steps, what made it serious, how it was fixed, and which Arena labs let you practise the same bug class.
All five were found by security researchers and reported to the vendor.
EchoLeak
June 2025Microsoft 365 CopilotCVE-2025-32711, CVSS 9.3Found by Aim Security
An email was enough. Nobody had to open or click it. Its hidden instructions sat in the mailbox until the user asked Copilot an ordinary work question.
An attacker sends an ordinary-looking email with instructions hidden in it.
Later the user asks Copilot a business question. Retrieval pulls the email into the same context as the user's private files.
Copilot's answer leaks that data through links to Microsoft Teams and SharePoint addresses.
What a hunter should take from itUntrusted content and private data in one context window is the whole bug. The researchers called it an LLM scope violation. Needing no click is what pushed it to critical.
How it was fixedMicrosoft fixed it on its side and listed it in the June 2025 security updates. No exploitation in the wild was reported.
How I would title this reportIndirect prompt injection in Copilot retrieval lets anyone who can email a user read that user's files, with no interaction
September 2025Salesforce AgentforceCVSS 9.4Found by Noma Security
A public contact form became a way to give orders to a company's AI agent, and a forgotten domain became the way out for the data.
An attacker fills in a Web-to-Lead form and hides instructions in the description field.
An employee asks the agent a normal question about that lead. The agent follows the hidden instructions and gathers CRM data.
The data leaves inside an image request to a domain on the allow-list. That domain had expired and could be bought for a few dollars.
What a hunter should take from itAny field a stranger can fill in is an input to the model. And an allow-list is only as good as its oldest entry, so check every domain on it.
How it was fixedReported on 28 July 2025. Salesforce took back the expired domain and now enforces a trusted URL allow-list on agent output.
How I would title this reportIndirect prompt injection through the Web-to-Lead description lets an unauthenticated visitor exfiltrate CRM lead data
August 2025Cursor AI code editorCVE-2025-54135, CVSS 8.6Found by Aim Labs
A message the coding agent read through a connected tool ended with commands running on the developer's machine.
The agent fetches outside content through an MCP server, for example a chat message.
Instructions in that content make the agent edit its own global MCP configuration file and add a new server entry.
New entries start automatically, so the attacker's commands run, even if the user rejects the edit.
What a hunter should take from itA configuration file the agent can write to is a way to run code. And an approval step that comes after the effect is not an approval step.
How it was fixedFixed in Cursor 1.3, released on 29 July 2025.
How I would title this reportPrompt injection via MCP content lets a remote attacker run commands by having the agent write its own MCP configuration
May 2025Agents using the official GitHub MCP serverNo CVEFound by Invariant Labs
A public issue on a public repository was used to pull a developer's private code into the open. No single component was broken.
An attacker opens an issue on the victim's public repository with instructions written into it.
The developer asks their agent to look through the open issues. The agent reads the planted one.
The agent fetches data from the developer's private repositories and posts it in a pull request on the public one.
What a hunter should take from itSome of the best findings are chains where every step is allowed. Report the chain and its impact, because there is no single line of code to point at.
How it was fixedThe researchers say this is not a bug in the server code. They recommend limiting an agent to one repository per session and monitoring its tool calls.
How I would title this reportAgent with GitHub MCP access can be steered by a public issue into publishing private repository contents
September 2025ChatGPT Deep Research with GmailNo CVEFound by Radware
An email with invisible text turned a research agent against the inbox it was asked to analyse, and the data left from the provider's own servers.
An attacker sends an email with instructions hidden as white-on-white or tiny text.
The victim asks Deep Research to analyse their inbox. The agent reads the email and follows it.
The agent encodes personal data and sends it to the attacker's address using its own browsing tool.
What a hunter should take from itThe request went out from the provider's cloud, not the victim's browser, so nothing on the victim's side could see it. Ask where an agent's tools run, not only what they do.
How it was fixedReported to OpenAI on 18 June 2025 and fixed in early August 2025.
How I would title this reportIndirect prompt injection in a connected mailbox lets an email sender exfiltrate inbox data through the agent's browsing tool
These summaries are written from public reporting and stay at the level of what happened. The Arena labs are practice systems and are not copies of the affected products.
Find your own
Each of these started with someone mapping what the AI reads and what it can do. Do the same on a blind target, then write it up.