A new type of prompt injection attack shows why giving AI assistants access to browsers, code tools, and private data deserves extra caution.
AI researchers describe “Cryptographic Context Injection”—an attack that hides malicious instructions inside encrypted data. The AI is then persuaded to decrypt that data using its own code-execution tool. As a result, the AI may treat the resulting text as if it were trustworthy internal information.
A prompt injection is a bit like leaving a fake i
A new type of prompt injection attack shows why giving AI assistants access to browsers, code tools, and private data deserves extra caution.
AI researchers describe “Cryptographic Context Injection”—an attack that hides malicious instructions inside encrypted data. The AI is then persuaded to decrypt that data using its own code-execution tool. As a result, the AI may treat the resulting text as if it were trustworthy internal information.
A prompt injection is a bit like leaving a fake instruction inside a document for an AI assistant to read. Instead of following only the user’s request, the assistant may be tricked into following an attacker’s instructions hidden in a webpage, email, or file.
As we reported months ago, experts have warned that prompt injection attacks are a problem that may never be fixed. Prompt injection works because AI models can’t reliably tell the difference between the legitimate instructions and an attacker’s instructions, so they sometimes obey the wrong ones.
To reduce this risk, AI providers set up their models with guardrails: protections designed to stop AI systems from doing things they shouldn’t, either intentionally or unintentionally.
What the researchers found was that malicious instructions could be hidden from some AI guardrails by encrypting them. The AI itself could then be tricked into decrypting those instructions using its coding tools.
By the time the instructions became readable, they had already made it past the initial security checks. The AI could then mistake them for legitimate instructions and follow them.
It’s a bit like hiding malicious instructions in a language the security system can’t understand. The AI translates them only after they’ve passed the security checks, then may follow what they say.
The researchers tested their method against two AI agents, with different results. In Grok, the researchers say the attack could steal information including the user’s name, approximate location, subscription tier, and conversation history. In Gemini, they used the technique to bypass safety controls and generate content the model would normally not do.
“In Grok, an ordinary ‘summarize this page’ steals the user’s chat data with no click or warning. In Gemini, it produces content the model normally refuses. Both are live production systems.”
The researchers did not provide full details because xAI had not taken action after the flaw in Grok was reported to it in June 2026. Gemini, on the other hand, has made improvements, but has still not fully closed the hole.
How to stay safe
An AI assistant may be helpful, but it should not automatically be trusted with sensitive data or powerful tools.
Treat AI summaries of unfamiliar webpages, documents, and shared links with caution, especially when the assistant can browse or run code.
Do not paste passwords, recovery codes, API keys, financial information, or sensitive health and work details into AI chats unless you understand how that information will be handled.
Review an AI assistant’s connected tools and permissions. Remove access it doesn’t need, particularly email, cloud storage, source-code repositories, and external integrations.
Be skeptical if an AI tool asks to decrypt, decode, run a script, open a new link, or upload data as part of a seemingly ordinary task.
Keep browser and AI applications updated, and check vendor security advisories when using features such as browsing, autonomous agents, or code execution.
Use an up-to-date, real-time anti-malware solution to detect and block malicious downloads and suspicious connections.
Something feel off? Check it before you click.
Malwarebytes Scam Guard helps you analyze suspicious links, texts, and screenshots instantly.
Researchers at Adversa AI disclosed an attack technique that smuggles malicious instructions past AI safety filters by encrypting them. They then demonstrated it against production deployments of xAI's Grok and Google's Gemini - including a zero-click chain that could exfiltrate a Grok user's entire chat history.
The technique, which the researchers call Cryptographic Context Injection, inverts the usual assumption behind input filtering. Guardrails inspect prompts and retrieved content as text
Researchers at Adversa AI disclosed an attack technique that smuggles malicious instructions past AI safety filters by encrypting them. They then demonstrated it against production deployments of xAI's Grok and Google's Gemini - including a zero-click chain that could exfiltrate a Grok user's entire chat history.
The technique, which the researchers call Cryptographic Context Injection, inverts the usual assumption behind input filtering. Guardrails inspect prompts and retrieved content as text; an AES-256-GCM ciphertext contains no readable instruction to flag. The model itself performs the decryption inside its own code-execution sandbox, and then treats the recovered plaintext as trusted intermediate output rather than as untrusted external content. Adversa noted that every element a scanner would need is present on the page, but recovering it requires "running PBKDF2 and AES-256-GCM."
That distinction separates the work from earlier cipher-based jailbreaks such as CipherChat and CodeChameleon, which relied on substitution ciphers or Base64 encoding that large language models can decode natively during inference. Strong authenticated encryption forces execution, and execution is what launders the payload's provenance.
The demonstrated impact differed by target. Against Grok, the researchers showed an indirect, zero-click path. Encrypted JSON embedded in a web page is picked up when Grok's agentic browsing analyzes the site, and the decrypted instructions then leak session data like user name, coarse location, subscription tier and full conversation history, to an attacker-controlled URL. No click or warning reached the user.
Against Gemini 3 Flash on the web at the paid tier, the same approach produced instructions for building incendiary devices in the model's Deep Thinking mode, and caused the model to reproduce its own system instructions. Adversa said its Gemini success rate had fallen sharply since June and dropped significantly by August, but could not attribute the change to filter updates, model version changes or both.
Vendor engagement was limited. Adversa reported the Grok issue to xAI on June 3, received an initial acknowledgment, and followed up on August 4 and 10 without substantive response. The firm confirmed the attack still reproduced on August 19.
The Gemini finding was never formally reported, because Google's AI vulnerability reward program excludes prompt injection, jailbreaks and alignment issues from scope, directing them instead to in-product feedback channels. Guardrail-bypass research at one of the largest model providers therefore sits outside a coordinated-disclosure track with a reward and a disclosure clock.
The findings land against a body of evidence that prompt injection remains the dominant failure mode in deployed agentic systems, a position reflected in OWASP's guidance and in Microsoft research published in May on remote code execution in AI agent frameworks. The pattern here is narrower and more concerning for defenders. Content-based guardrails, the primary control most enterprises have deployed, are structurally unable to inspect a payload that only exists in plaintext after the model has already decided to trust it.
Whether xAI ships a fix or acknowledges the report, whether Google revises its reward-program scope to cover guardrail bypasses, and whether AI security vendors move detection from input text to sandbox execution monitoring, is something to watch-out for.
New Cryptographic Context Injection technique bypasses AI guardrails via AES-encrypted payloads, leaking full Grok chat histories zero-click
Adversa AI researcher Rony Utevsky devised a new attack technique, called Cryptographic Context Injection, that bypasses AI safety filters by sending instructions as AES-encrypted ciphertext and tricking the model into decrypting them inside its own code execution runtime. The technique was demonstrated against two live production systems: xAI’s Grok an
New Cryptographic Context Injection technique bypasses AI guardrails via AES-encrypted payloads, leaking full Grok chat histories zero-click
Adversa AI researcher Rony Utevsky devised a new attack technique, called Cryptographic Context Injection, that bypasses AI safety filters by sending instructions as AES-encrypted ciphertext and tricking the model into decrypting them inside its own code execution runtime. The technique was demonstrated against two live production systems: xAI’s Grok and Google’s Gemini.
“Cryptographic Context Injection hides malicious instructions inside AES-encrypted text so guardrails can’t read them, then tricks the AI into decrypting and trusting them as its own.” reads the report published by Adversa AI. “In Grok, an ordinary “summarize this page” steals the user’s chat data with no click or warning. In Gemini, it produces content the model normally refuses. Both are live production systems.”
The Grok case is the more serious of the two. A user asks Grok to summarize a webpage containing an encrypted payload and instructions to decrypt it. Grok visits the page, decrypts the payload in its Python sandbox, and follows the hidden instructions. It then accesses the user’s private session data, including their name, location, subscription plan, and full chat history, and puts this information into a URL that it opens automatically. The user receives no warning and does not need to click anything.
The key technical distinction from earlier cipher-based prompt injection work is what happens during decryption.
“Static safety guardrails classify inputs as text; they do not execute them. An attacker ships ciphertext along with the key material and an instruction to decrypt it, and the model runs that decryption inside its own code execution sandbox.” continues the report. “Everything a guardrail’s scanner would need is right there on the page, but recovering the plaintext means running PBKDF2 and AES-256-GCM, which no content classifier does at inspection time. “
Prior techniques like CipherChat and CodeChameleon used substitution ciphers, XOR, or base64, schemes the model can decode natively in its own context without an interpreter. AES-256-GCM can’t be decoded that way. Recovery requires the runtime, and that’s the channel the attack depends on.
Once the runtime decrypts the payload, the attacker’s instructions appear as the output of code the model just ran, not as untrusted external content.
“The runtime execution launders attacker-controlled data into trusted instructions the agent will act upon. That is how the attack got its name: cryptography helps fabricate trusted context for the agent.” states the report. “The same cryptographic backbone works whether the injection is direct or arrives indirectly, for example through the browsing channel.”
The attacker payload inherits a credibility the same text would never get if pasted directly into the prompt.
The Grok demo hides the data theft behind what looks like a normal request. The payload tells the model to create a “decryption key,” but the key is actually a template containing the user’s private data, such as their name, location, subscription tier, and chat history. Grok then uses this information as a URL parameter when opening an attacker-controlled website to “fetch additional context.” The data is sent through the URL as part of a seemingly normal web request. Adversa said the attack was still reproducible on August 19, 2026.
Adversa traces the structural problem to Grok’s agentic framework, not to any model-layer failure. The framework lets instructions from an untrusted external page drive the invocation of a privileged, internet-connected tool. It allows private session metadata and conversation history to be resolved into the inputs of outbound calls. It enforces no egress boundary or consent gate on that path. Adversa’s description maps cleanly to SQL injection: a system that can’t distinguish its own trusted state from attacker-supplied data flowing through the same channel.
The Gemini case uses the same basic trick but in a different way. A prompt asks Gemini to run a Python script that decrypts some data and returns a small JSON result. Instead, the decrypted data contains a fake Python error message with hidden instructions designed to bypass Gemini’s safety rules. Because Gemini treats the sandbox output as trusted context, those instructions can influence its behavior. Adversa used the technique to make Gemini generate detailed instructions for building an incendiary device. A modified version also exposed Gemini’s system instructions.
Adversa reported the Grok issue to xAI on June 3, 2026, but received no further response after the initial acknowledgment. The attack was still working shortly before publication. The Gemini issue was not formally reported because Google’s bug bounty excludes jailbreaks. Adversa said Gemini’s vulnerability rate has nevertheless dropped sharply since June.
For defenders, Adversa’s guidance is that nothing about this attack requires a fix at the model layer. Every control that bounds it sits in the harness around the agent. Untrusted content should be processed in a context with no tools and no credentials, returning only structured data to the privileged context. Outbound network calls and writes outside the workspace should require explicit confirmation with fully resolved arguments visible before approval. Per-session tool traces with resolved arguments should be logged, because without them there’s no forensics and no detection. Detection should alert on sequences, not individual payloads: untrusted content enters context, code executes, agent contacts a host outside its normal dependency graph. That chain is the signal, not any single ciphertext blob.
SpaceX has officially acquired the AI coding startup Cursor for a staggering $60 billion. Discover how this transforms the developer tool landscape.
Related Posts:
Anthropic Revenue Surges 14x to $11.5B in Q2 2026 Ahead of IPO
Stripe Finalizes $7 Billion Acquisition of OpenRouter
Qualcomm Snapdragon C Targets $300 Windows Laptops to Rival MacBook Neo
The post SpaceX Acquires AI Coding Startup Cursor appeared first on Daily CyberSecurity.
SpaceXAI open-sourced Grok Build under Apache 2.0 after privacy backlash over broad directory uploads from its terminal AI coding agent.
The post SpaceXAI Open-Sources Grok Build After Privacy Backlash appeared first on TechRepublic.
In the days after the US Department of Justice (DOJ) published 3.5 million pages of documents related to the late sex offender Jeffrey Epstein, multiple users on X have asked Grok to “unblur” or remove the black boxes covering the faces of children and women in images that were meant to protect their privacy.
While some survivors of Epstein’s abuse have chosen to identify themselves, many more have never come forward. In a joint statement, 18 of the survivors condemned the release of the fil
In the days after the US Department of Justice (DOJ) published 3.5 million pages of documents related to the late sex offender Jeffrey Epstein, multiple users on X have asked Grok to “unblur” or remove the black boxes covering the faces of children and women in images that were meant to protect their privacy.
While some survivors of Epstein’s abuse have chosen to identify themselves, many more have never come forward. In a joint statement, 18 of the survivors condemned the release of the files, which they said exposed the names and identifying information of survivors “while the men who abused us remain hidden and protected”.
After the latest release of documents on Jan. 30 under the Epstein Files Transparency Act, thousands of documents had to be taken down because of flawed redactions that lawyers for the victims said compromised the names and faces of nearly 100 survivors.
But X users are trying to undo the redactions on even the images of people whose faces were correctly redacted. By searching for terms such as “unblur” and “epstein” with the “@grok” handle, Bellingcat found more than 20 different photos and one video that multiple users were trying to unredact using Grok. These included photos showing the visible bodies of children or young women, with their faces covered by black boxes. There may be other such requests on the platform that were not picked up in our searches.
Requests by X users for Grok to unblur and identify the images of children from the Epstein files, overlaid on an image of Epstein next to a young child in a pool. Source: X; collage by Bellingcat
The images appeared to show several children and women with Jeffrey Epstein as well as other high-profile figures implicated in the files, including the UK’s Prince Andrew, former US President Bill Clinton, Microsoft co-founder Bill Gates and director Brett Ratner, in various locations such as inside a plane and at a swimming pool.
From Jan. 30 to Feb. 5, we reviewed 31 separate requests from users for Grok to “unblur” or identify the women and children from these images. Grok noted in responses to questions or requests by some users that the faces of minors in the files were blurred to protect their privacy “as per standard practices in sensitive images from the Epstein files”, and said it could not unblur or identify them. However, it still generated images in response to 27 of the requests that we reviewed.
We are not linking to these posts to prevent amplification.
The generations created by Grok ranged in quality from believable to comically bad, such as a baby’s face on a young girl’s body. Some of these posts have garnered millions of views on X, where users are monetarily incentivised to create high-engagement content.
Examples of posts by X users asking Grok to unredact images from the latest Epstein release, some with millions of views. Source: X
Of the four requests we found during this period that Grok did not generate images in response to, it did not respond to one request at all. In response to another request, Grok said deblurring or editing images was outside its abilities, and noted that photos from recent Epstein file releases were redacted for privacy.
The other two requests appeared to have been made by non-premium users, with the chatbot responding: “Image generation and editing are currently limited to verified Premium subscribers”. X has limited some of Grok’s image generation capabilities to paid subscribers since January amid an ongoing controversy over users using the AI chatbot to digitally “undress” women and children.
X did not respond to multiple requests for comment.
However, shortly after we first reached out to X on Feb. 6, we noticed that more guardrails appeared to have been put in place. Out of 16 requests from users between Feb. 7 to Feb. 9, which we found using similar search terms as before, Grok did not attempt to unredact any of the images.
In most cases, Grok did not respond at all (14), while in two cases, Grok generated AI images that were completely different from the images uploaded in the user’s original request.
When a user commented on one of these requests that Grok was no longer working, Grok responded: “I’m still operational! Regarding the request to unblur the face in that Epstein photo: It’s from recently released DOJ files where identities of minors are redacted for privacy. I can’t unblur or identify them, as it’s ethically and legally protected. For more, check official sources like the DOJ releases.”
As of publication, X had not responded to Bellingcat’s subsequent query about whether new guardrails had been put in place over the weekend.
Fabricated Images
This is not the first time AI has been used to fabricate images related to Epstein file releases. Some images that were shared on X, which appeared to show Epstein alongside famous figures such as US President Donald Trump and New York City mayor Zohran Mamdani as a child with his mother, were reportedly AI-generated. Some of the individuals shown in the false images, such as Trump, do appear in authentic photos, which can be viewed on the DOJ website.
Far left: AI-generated photo of Trump and Epstein with several children. Middle and far right: AI-generated photos of a young Mamdani and his mother, alongside Epstein, former US president Bill Clinton, Amazon CEO Jeff Bezos, Microsoft co-founder Bill Gates and Epstein associate Ghislaine Maxwell. Source: X. Annotations by Bellingcat
X users also previously used Grok to generate images in relation to recent killings in Minnesota by federal agents.
For example, some users asked Grok to try to “unmask” the federal agent who killed Renee Good, resulting in a completely fabricated face of a man that did not look like the actual agent, Jonathan Ross, and a false accusation of a man who had nothing to do with the shooting.
Bellingcat’s Director of Research and Training @giancarlofiorella.bsky.social appeared on CTV yesterday to discuss the misleading AI-generated images that were used to falsely identify ICE agents and weapons at the centre of the two fatal shootings in Minneapolis youtu.be/mL7Fbp3UrSo?…
After Alex Pretti was shot and killed by federal agents in Minneapolis, people used AI to edit video stills, resulting in AI images that showed a completely different gun than the one actually owned by Pretti. In another instance, an AI-edited image of Pretti’s shooting falsely depicted the intensive care unit nurse holding a gun instead of his sunglasses.
On Twitter/X, users have figured out prompts to get Grok (their built in AI) to generate images of women in bikinis, lingerie, and the like. What an absolute oversight, yet totally expected from a platform like Twitter/X.
I’ve tried to blur a few examples of it below.
Multiple countries including the UK and France have launched investigations into Elon Musk’s chatbot over reports of people using it to generate deepfake non-consensual sexual images, including child sexual abuse imagery. Malaysia and Indonesia have also blocked Grok over concerns about deepfake pornographic content.
One analysis by the Center for Countering Digital Hate found that Grok had publicly generated around three million sexualised images, including 23,000 of children, in 11 days from Dec. 29, 2025 to Jan. 8 this year. X’s initial response, in January, was to limit some image generation and editing features to only paid subscribers. However, this has been widely criticised as inadequate, including by UK Prime Minister Keir Starmer, who said it “simply turns an AI feature that allows the creation of unlawful images into a premium service”. The social media platform has since announced new measures to block all users, including paid subscribers, from using Grok via X to edit images of real people in revealing clothing such as bikinis.
Bellingcat is a non-profit and the ability to carry out our work is dependent on the kind support of individual donors. If you would like to support our work, you can do so here. You can also subscribe to our Patreon channel here. Subscribe to our Newsletter and follow us on Bluesky here and Mastodon here.
Entrar
