Thursday, November 19, 2026 Breach Index $11.1B lost to breaches in 2026 · 117 breaches made public
How attacks workAI

Copilot, Grok, isolated-vm – One Bug Behind All Three

Chat becoming command on AI.

Between August 18th and August 20th, three separate AI security research findings went public. Microsoft fixed a flaw in Copilot that could empty your Gmail by means of a single link. A research lab published an attack that slips straight past safety filters by encrypting itself. And researchers found a flaw in the foundation of a sandbox library that much of the AI agent industry relies on.

Three companies, three different layers of the stack, three research teams working independently. Taken one at a time, they sounded like a bad week.

It isn’t a bad week. It’s the same issue showing up in three places, and once you see it, you can’t stop seeing it.

What actually happened

Copilot told researchers how to hack it

Varonis Threat Labs found a chain of flaws in Microsoft Copilot Personal, now tracked as CVE-2026-24301 and named CoSnitch. Microsoft released a patch on August 18.

The chain has three links. It starts with execution. The undocumented URL parameter autorun=1, combined with Copilot’s standard ?q= query parameter, runs the attacker’s prompt as soon as the page loads. No click, no confirmation, and nothing on screen to show that anything happened.

From there, the prompt could reach anything the victim had connected: Gmail, Drive, Calendar, OneDrive, and Copilot’s own chat history. It packaged the results into a URL and sent them, through Copilot’s built-in fetch function, to a server the attacker controlled. Varonis noted that, at the network layer, the outgoing request looks identical to Copilot fetching a page it was asked to summarise. So your security tools see a normal AI assistant doing normal AI assistant things.

The third link is the one that should worry you most. If Copilot summarises a specially crafted web page, the attacker’s instructions can end up in Copilot’s permanent memory. Not the session. The memory. The injection survives a password change, session revocation, and even a device wipe followed by enrolling a new one. It stays there until someone opens Copilot’s memory settings and deletes it by hand, and the whole time it creates no process, no file, and no log entry that would trip an alarm.

As for how they found it: the researchers kept asking Copilot why their attacks weren’t working. Copilot, being helpful, explained. In doing so, it revealed enough about its own architecture for the researchers to identify the undocumented parameter they needed. Varonis calls this meta-hacking. It’s social engineering aimed at the model’s reasoning instead of a person’s trust.

One more thing worth knowing: Varonis reported the issue on December 31, 2025. Microsoft fixed part of it in February and the rest on August 18. That’s almost eight months, and any memory poisoning planted during that window is still sitting in people’s accounts.

An attack that encrypts itself past the guardrails

Adversa AI has published a technique it calls Cryptographic Context Injection, and the core idea is almost annoyingly simple.

Safety guardrails read text and classify it. They look at the words and decide whether those words describe something harmful. So Adversa stopped sending words. It sent ciphertext instead, along with the key material and an instruction telling the model to decrypt it.

The guardrail sees gibberish. It can’t flag that gibberish as dangerous, because recovering the real instruction means running PBKDF2 and AES-256-GCM, and content classifiers don’t do cryptography. So it lets the payload through. The model, however, has a code execution sandbox. It decrypts the payload right there in its own environment and then does whatever the payload says.

Against Grok, the attack leaked the victim’s entire chat history without a single click. Everything the guardrail needed to catch it was right there on the page. It just wasn’t readable at the moment of the check.

The disclosure timeline is a story of its own. Adversa notified xAI on June 3, 2026, via the company’s HackerOne programme. xAI acknowledged the report but gave no further details and no timeline. Adversa followed up twice, on August 4 and August 10, and heard nothing back. On August 20, Adversa published its findings. The attack still worked the day before, and Adversa deliberately left out the working payloads. The same method worked against Gemini in June, though its success rate had fallen considerably by August. Adversa did not know why, and Google never received a formal report because its disclosure programme does not cover jailbreaks.

One of these is patched. The other is a live zero-day as you read this.

The sandbox everyone builds on has a hole

On August 19, Endor Labs disclosed a serious vulnerability in isolated-vm, tracked as GHSA-864f-rcv7-6rh4. No CVE has been assigned yet.

Even if you’ve never heard of isolated-vm, you’ve probably used a tool built on it. It’s the Node.js library that runs untrusted JavaScript inside a separate V8 isolate, and in practice it’s how a large share of AI agent platforms let a model execute code without giving that code access to the host system. It gets about a million npm downloads a week.

The isolate itself remained unscathed. What failed was the C++ code that passes values across the boundary: specifically, a type confusion in how ExternalCopy handles the transferList option. As the researchers put it, they didn’t break the sandbox. They broke the code that carries data into it. Starting from a single reference object, which is the normal way a host grants a sandbox any capability, they went from a crash to hijacking the host process’s control flow. That’s a full escape from guest to host.

Versions up to and including 7.0.0 are affected. The fixes shipped in 6.2.0 and 7.0.1.

This differs from a typical remote code execution bug in one key way. Normally, an attacker has to find a way to reach the vulnerable code. On an agent platform, the entire product is a machine that accepts attacker-influenced code and runs it. There’s no precondition to clear, because running untrusted code is the feature.

The pattern

Look at exactly what went wrong in each case.

Copilot took in a webpage as data, then followed its contents as instructions. The guardrail let ciphertext through as data, and the model turned it into instructions. isolated-vm accepted a value as data, and it came out as control flow in the host process.

Same failure. Three altitudes.

Each system lost track of the line between ‘the stuff I’m processing’ and ‘the stuff I’m obeying’. Varonis’s Lior Adar called this a fundamental architectural problem in LLMs, not a flaw in any one product.

Traditional software dealt with this problem the hard way, mostly by accident. SQL injection is the same bug: user input ends up in the query. Cross-site scripting is the same bug too: user input ends up on the page. Parameterised queries and output encoding fixed them, because a database query has a real, structural line between code and data that can be enforced.

A language model has no such boundary. The system prompt, the user’s question, the webpage it just retrieved, and the tool result it got back all arrive as tokens in the same context window. Nothing in the syntax marks which part is authoritative and which is just content. The model has to infer the difference, and that inference can be contested.

That’s exactly why these incidents keep happening, and why every fix so far has bolted guardrails on from the outside instead of changing the structure. Adversa’s attack shows what a bolted-on guardrail is worth: encrypt the payload, and the classifier has nothing to classify.

Why the usual advice fails here

After a credential-related breach, the standard playbook is to rotate passwords, kill sessions, and re-enroll devices. All of that fails against CoSnitch’s memory poisoning, because the attacker’s instructions weren’t in a session or a credential. They were in the assistant’s memory of you.

It’s a completely new kind of problem. For years we’ve taught people that changing a password fixes a compromised account. It doesn’t fix this. Nothing in most teams’ incident response habits points you toward “checking the AI assistant’s stored memories”, and no log tells you to.

The isolated-vm bug undermines a different assumption. Many AI platform architectures rest on one premise: it’s safe to run untrusted code because it executes in a sandbox. The compliance argument, the risk assessment, and the reason nobody built a second layer all depend on that premise holding. Once it breaks, nothing else is there to catch the failure, because the whole idea of the design was that nothing else was needed.

If you use AI assistants

In practice, roughly in order of importance.

Check what your assistant has saved to memory. In Copilot, ChatGPT, Gemini, or anywhere else, look at what’s actually stored there. If you find anything that looks like an instruction you never gave, delete it. It takes two minutes, and almost nobody has done it.

Be more careful with connected apps. How much damage CoSnitch could do depended entirely on which apps the victim had connected. An assistant with Gmail, Drive and Calendar connected can read all three for whoever manages to send it a prompt. Connect only the apps you need, and disconnect the ones you hooked up for a task in March and never removed. The same basic security hygiene that applies everywhere else applies here.

Treat a link that opens an AI tool with the same suspicion as any other link. There’s now a kind of URL that loads a chat interface with a prompt already filled in.

Be suspicious of summaries. When you ask your assistant to summarise a page, you’re effectively feeding the contents of that page into your session. Most pages are fine. The attack depends on the ones that aren’t.

If you build with them

Adversa’s mitigation advice is the most useful thing published this week, and it applies well beyond the specific attack it describes. Instead of trying to detect malicious payloads, focus on provenance and tag where content came from. Block any tool call whose arguments derive from fetched or decrypted content. Alert on the whole sequence of actions rather than on any single input, because the single input is exactly what an attacker can make unreadable.

For the sandbox bug, the obvious step is to patch to 6.2.0 or 7.0.1. The bigger lesson: if your threat model says “contained by sandbox”, it needs a second line. Not because isolated-vm is bad. The researchers were clear that it’s still the best option available. It’s because a single control with no defence behind it is a single point of failure, however good it is.

The part that should bother you

Attackers didn’t find any of these three. Varonis, Adversa and Endor Labs are all defensive companies, and they published their work.

That’s the good news, and also the uncomfortable part, because the pattern they all exposed can’t be patched. Microsoft can fix autorun=1, xAI will eventually handle encrypted prompts, and isolated-vm has shipped a version without the type confusion. The underlying problem, that these systems can’t reliably tell content from commands, comes from how they’re built, not from a flaw in how they were built this time.

We’re wiring assistants into email, calendars, files and code execution at speed, as if the boundary problem were already solved. It isn’t. It’s being handled case by case, by whichever researchers get there first.

This week, they did. That’s worth something. It isn’t the same as being safe.


Sources

Found this useful? Save it for later or send it to someone who needs it.

Written by

Hubert Ramsey

See all 5 articles

Up next

What is a brute force attack?

What is a brute force attack? A brute-force attack guesses the password; here, the attacker runs software to try millions of passwords, usernames, and keys until…

Keep reading · 6 min read

Join the discussion

Your email won't be published.