What is the Whisper Leak attack on AI chatbots?
Whisper Leak infers AI chatbot topics from encrypted traffic. See the measured results, vendor fixes, and limits.

Whisper Leak is a side-channel attack on AI chatbots. Microsoft researchers published it in November 2025. It infers a chat's topic from encrypted traffic. It uses packet sizes and timing, not the text. The attacker must be able to watch your traffic. That includes ISPs and people on shared Wi-Fi. The attack does not break TLS encryption.
The main limit is scope. The attack flags whether a chat matches one trained topic. The paper tested one: the legality of money laundering. Four providers have since added mitigations.
What is Whisper Leak?
Whisper Leak comes from a Microsoft research paper. Geoff McDonald and Jonathan Bar Or wrote it. It appeared on arXiv as 2511.03675 in November 2025.
The attack does not break encryption. It reads how encrypted data flows instead. Chatbots generate replies in small units called tokens. A token is roughly a word or word fragment. Many services stream each token as it is generated. Each reply leaves a pattern of packet sizes and timing. A classifier can link that pattern to a topic.
Think of a sealed letter. You cannot read it. Its weight, size, and send time still tell you something. Whisper Leak applies that idea to AI chat traffic.
The research: 28 models tested for topic leaks
The researchers built a proof of concept. They tested it against 28 models from major providers:
- OpenAI's GPT-4o mini, GPT-4.1 mini and nano, and o1-mini
- Anthropic's Claude 3 Haiku
- Google's Gemini 2.5 Pro and Gemini 1.5 Flash models
- DeepSeek R1 and V3, plus DeepSeek R1 hosted by Microsoft
- xAI's Grok 2 and Grok 3 Mini
- Mistral Large and Mistral Small
- Alibaba's Qwen2.5 Plus and Turbo
- Amazon Nova Pro and Nova Lite
- Microsoft-hosted GPT and Phi models
- Llama models hosted on Groq and Lambda
The paper scores attacks with AUPRC, not plain accuracy. AUPRC means area under the precision-recall curve. In Table 1, 13 models scored 98% or higher. Four scored 99.9% or higher. The paper's summary text says 17 models passed 98%. Its own table does not match that.
The authors also projected a harder case. One target conversation hides among 10,000 others. There, 17 models reached 100% precision. They caught 5% to 20% of target conversations. The authors call this an extrapolation from test data. Their only test topic was the legality of money laundering.
An attacker does not see your messages. Traffic metadata alone can point to the topic.
How does the Whisper Leak attack work?
The attack targets streaming. Chatbots send replies while the model generates them. That creates the familiar typing effect. The attack has three steps.
1. Token-by-token streaming creates patterns
Different prompts produce different replies. Ask about money laundering, then about weather. The two replies differ in length and rhythm.
Many services send tokens the moment they are generated. Table 2 of the paper estimates tokens per network event. For 21 of 28 models, the estimate is about 1. Google, Anthropic, and Alibaba batch several tokens per event. TLS hides each record's content. Its size and timing stay visible.
2. Traffic analysis points to the topic
An observer on your network path cannot read encrypted packets. Think shared Wi-Fi, an ISP, or a network admin. The observer can still see three things:
- Packet sizes: How many bytes in each encrypted packet
- Timing intervals: How long between each packet
- Sequence patterns: The overall rhythm and flow of the conversation
Together they form a pattern. Different topics tend to leave different patterns.
3. Machine learning matches patterns
The researchers trained classifiers on recorded traffic. Each classifier answers one question. Is this chat about the target topic? A real attacker would need similar training data.
Which topics might an attacker train for? Microsoft's write-up names topics oppressive governments may target. Its examples are protests, banned material, elections, and journalism. The paper itself tested one topic.
None of this requires decrypting a packet.
Which AI models were most vulnerable in the tests?
Results varied widely by model.
High-risk models (highest AUPRC)
These 13 models scored 98% AUPRC or higher:
- Mistral: Mistral Large (100.0% AUPRC) and Mistral Small (99.9%)
- DeepSeek: R1 hosted by Microsoft (99.9%), R1 direct (99.4%), and V3 (98.3%)
- Groq: Llama 4 Maverick (99.7%)
- xAI: Grok 3 Mini (99.9%) and Grok 2 (99.0%)
- Alibaba: Qwen2.5 Plus (99.1%) and Turbo (99.0%)
- OpenAI: o1-mini (98.9%), GPT-4o mini (98.6%), and GPT-4.1 mini (98.0%)
Per Table 2, 11 streamed about 1 token per event. The two Qwen models averaged about 4.4 tokens.
More resistant (but not immune)
These models scored lowest in Table 1:
- Google Gemini: 81.9% to 84.0% AUPRC. The paper names Google among providers that batch tokens.
- Amazon Nova: 71.2% to 77.5% AUPRC
- Microsoft-hosted Phi-3.5 MoE: 76.9% AUPRC
The paper offers a tentative reason for Gemini and Nova. Architecture differences or heavier batching may explain it. It gives none for Phi.
"More resistant" does not mean safe. The attack still worked, with lower AUPRC.
Who can perform this attack?
The paper describes a passive network observer. It can watch encrypted traffic but cannot decrypt it. The paper names ISPs, government agencies, and local network observers. Microsoft adds someone connected to the same Wi-Fi router. The attacker also needs training data for the target topic.
Public Wi-Fi networks
Public Wi-Fi puts strangers on your network. The paper's own example is coffee shop Wi-Fi. Whoever runs the router can watch your traffic. A compromised router gives an attacker the same view.
Internet service providers (ISPs)
Your ISP carries your traffic to the internet. It can see the size and timing of encrypted packets. The paper names ISPs as one possible observer.
Corporate network administrators
On a company network, IT staff can observe traffic patterns. That includes the metadata of encrypted AI chats.
Compromised network equipment
Network gear along your path also sees your traffic. A compromised router or switch can collect the same metadata.
The attack is passive. It sends nothing to your device. It breaks no encryption. It only observes traffic the observer can already see.
Hypothetical scenarios: when a chat topic matters
These scenarios are illustrations, not documented incidents. The paper tested one topic only. A real attacker would need a classifier for each topic.
Scenario 1: an executive strategy session
A CFO travels to a conference. The CFO uses a chatbot to plan merger talking points. The hotel Wi-Fi carries the chat. The chat covers:
- Acquisition targets and valuation models
- Integration strategy and timeline
- Potential job cuts and restructuring plans
Suppose a spy on that network has a merger classifier. The traffic alone could flag the chat as merger-related. The spy would not learn which companies. That alone could prompt closer surveillance.
Possible impact: targeted surveillance or trading ahead of the news.
Scenario 2: a nurse researching a rare condition
A nurse asks a chatbot about a rare condition. The chat runs over the hospital network. The chat covers:
- Symptoms and diagnostic criteria
- Treatment protocols and drug interactions
- Prognosis and complications
A trained classifier could flag the chats as medical. That might hint at an unusual case. It would not reveal the patient.
Possible impact: patient privacy concerns and targeted social engineering.
Scenario 3: a researcher on hotel Wi-Fi
A drug researcher works on hotel Wi-Fi. The researcher asks a chatbot about formulations and patents.
A spy with the right classifier could flag the topic. The spy would not see the compounds. The flag could still single out the researcher.
Possible impact: targeted phishing or further espionage against the team.
Why this matters for browser security
AI chat often runs in a browser tab. Microsoft calls browsers the primary attack surface on client devices.
The browser is where you:
- Access AI chatbots (ChatGPT, Claude, Gemini)
- Handle sensitive business communications
- Process financial transactions
- Store authentication credentials
- Maintain your digital identity
Encryption is easy to mistake for privacy. Whisper Leak shows the gap. Encryption hides the words, not the traffic's shape.
The limits of encryption alone
Encryption protects message content. It does not hide:
- When you're communicating
- With which services
- How much data you're exchanging
- The patterns of your interactions
Whisper Leak exploits that metadata. The problem is older than AI chat. A 2017 USENIX Security paper studied encrypted video. Many streams had unique burst patterns.
Zero trust and metadata
Whisper Leak fits a core zero trust rule. Trust nothing, verify everything, and assume breach.
Encryption alone does not protect sensitive work. A network you use may not be secure. Traffic patterns can reveal activity.
How remote isolation changes the observer
Whisper Leak needs a view of the AI response stream. Remote isolation changes that path. It does not remove every observer.
Remote isolation loads the AI site in a browser elsewhere. Locally, an observer sees the isolation service connection. It no longer sees the AI service directly. That can take the token stream off the local link.
What still remains visible
The isolation connection still has timing and volume. The isolation operator can observe its own service. The AI provider still receives the conversation. Signing in still identifies the account.
- Local observer: sees the isolation-service connection and its traffic shape.
- Isolation operator: becomes another trust point in the path.
- AI provider: still sees the prompt, account, and service-side data.
- Signed-in identity: remains linkable regardless of browser location.
The honest conclusion is narrower. Isolation can change which observer sees the AI stream. Only testing shows whether it defeats a given classifier.
What AI providers have done so far
The authors began responsible disclosure in June 2025. Four providers deployed fixes before publication in November 2025. Dates come from the paper. Methods come from Microsoft's blog.
Provider fixes
OpenAI (August 2025): Added random-length obfuscation text. This adds noise to the traffic patterns, making fingerprinting harder.
Microsoft Azure (October 2025): Applied a similar obfuscation field.
Mistral (September 2025): Added a parameter with similar effect.
xAI (August 2025): Addressed the issue. The paper does not describe its method.
Limits of the tested mitigations
The paper tested random padding, token batching, and packet injection. Each cut AUPRC by 3.5 to 4.8 points. They don't eliminate the vulnerability. Microsoft's blog is more optimistic about Azure's deployed fix. Microsoft no longer considers the risk practical there. Caution still makes sense:
- Residual signal: Every mitigation the paper tested left measurable signal behind.
- Performance tradeoffs: The paper says padding adds bandwidth and latency costs. Packet injection roughly doubles or triples traffic.
- Incomplete adoption: Not all providers have implemented mitigations. Some declined to act or didn't respond to the disclosure.
- Arms race dynamics: As padding techniques evolve, so will attack methods. The paper calls this a likely cat-and-mouse game.
The core issue remains. Where others can watch the network, metadata can leak.
Practical steps for enterprise AI use
Does your organization use AI chatbots for sensitive work? Start with these steps.
Immediate actions
1. Audit AI use across your organization
- Identify which teams are using which AI services
- Catalog what types of information are being discussed
- Assess the sensitivity of these conversations
2. Segment network traffic
- Separate AI traffic from general internet usage where possible
- Use dedicated, monitored channels for sensitive AI interactions
- Implement VPN requirements for remote workers accessing AI tools
3. Update security policies
- Set rules for what staff may share with chatbots
- Require approval processes for sensitive use cases
- Train employees on metadata risks, not just content risks
Longer-term architecture
4. Evaluate isolation-based controls
- Map which observers can see each network segment
- Test the exact routing or isolation implementation
- Document which service becomes the new trust point
5. Apply zero trust principles
- Assume all networks are untrusted, including your own
- Verify all activities, not just authenticate users
- Monitor for unusual patterns even in encrypted traffic
6. Plan for vendor diversity
- Don't rely on a single AI provider's mitigation efforts
- Use multiple models for different sensitivity levels
- Prefer models with token batching where appropriate
The broader lesson: encryption is not enough
Whisper Leak is not the first traffic analysis attack. It shows a familiar gap. What seems protected is not always protected.
The lock icon is easy to misread. HTTPS hides what you send. It does not hide when, where, or how much.
Traffic analysis targets what encryption leaves exposed:
- Timing attacks: Identifying activities by when they occur
- Traffic volume analysis: Inferring activity from data volumes
- Behavioral fingerprinting: Recognizing patterns in how you interact with services
- Metadata correlation: Linking encrypted activities across services and time
These attacks use how networks work, not a software bug. That makes them hard to patch away.
Stronger encryption alone will not fix this. Rethink where sensitive work runs, and who can watch.
Conclusion: map each observer
Encryption alone does not hide a chat's topic. Whisper Leak shows that clearly. AI chatbots are now common work tools. Security plans should cover metadata, not just content.
Encryption protects message content. It does not protect the metadata around it. Metadata can show what you do and when.
The question is not whether to use AI tools. It is how to use them without new exposure.
Browser routing and isolation change the observer's vantage point. Legba is a Chrome extension with two modes. Ghost gives you a private browser route. Shield opens a page in an isolated browser off your device. These controls reduce some exposure. They do not erase every traffic signal. Read the architecture in how Legba browser isolation works.
Encrypted chats can still expose traffic patterns. The goal is to limit what each observer learns.
No routing or isolation layer makes traffic analysis impossible. Treat it as one control inside a broader privacy model.
Sources were checked on October 30 and November 1, 2026.
Continue the security analysis
Related reading on AI privacy, browser isolation, and traffic exposure.
What is shadow AI? Why the browser tab matters
Shadow AI is unauthorized workplace use of AI tools. See the risks, governance gaps, and where browser isolation changes one boundary.
How Legba browser isolation works
Shield runs the page in an isolated browser. That browser is off your device. See what happens when you open the tab. See what ends when you close it.
Browser isolation vs VPN: which tool fits in 2026?
VPNs protect a network path. Browser isolation changes where web content runs. Compare their scope, limits, and best uses.
Free for 30 days. No card required.
Choose where the page runs
Legba is a Chrome extension with two modes. Ghost gives you a private browser route. Shield opens a page in an isolated browser off your device.
Frequently asked questions
Can I detect if someone is performing a Whisper Leak attack on my traffic?
Not from your device. The paper's attacker is passive. It records packets and sends nothing to your device. The attack itself gives your device nothing to flag. Limit who can observe your traffic instead.
Does using a VPN protect against Whisper Leak?
Only in part. A VPN hides your destination from local observers. They still see the tunnel's timing and volume. Your VPN provider can still observe the traffic. Observers past the VPN server could still run the attack. Microsoft's write-up suggests a VPN as an added layer. VPNs shift the trust point but don't eliminate the vulnerability.
Are on-device AI models (like running Llama locally) immune to Whisper Leak?
Yes, if the model runs entirely on your device. With no network traffic, there is no stream to observe. Check whether your local app sends data elsewhere.
Which AI chatbots should I use for sensitive conversations?
Three model families scored lowest in the paper. They are Amazon Nova, Microsoft-hosted Phi-3.5, and Google Gemini. None of them is immune. The paper names Google among providers that batch tokens. Microsoft suggests providers with mitigations and non-streaming models. Share secrets only where policy and your employer allow. Routing and isolation leave the provider's access unchanged.
How can I tell if my AI service has implemented Whisper Leak mitigations?
Four providers deployed fixes from August to October 2025. The paper names OpenAI, xAI, Mistral, and Microsoft. Other providers declined or did not respond. Tested mitigations reduced the attack but did not end it. Check your AI provider's security bulletins for official statements.
Does Whisper Leak affect other streaming services like Netflix or Spotify?
Whisper Leak studies streamed AI chat replies. The paper did not test video or music services. Encrypted video has its own known side channel. A 2017 USENIX Security paper studied it. It identified video streams by their burst patterns.
What should I do if I've already had sensitive conversations with AI chatbots?
Past traffic cannot be protected after the fact. If someone was watching then, that capture already exists. You can still act:
- Assume those conversation topics may have been identified
- Implement stronger security measures going forward
- Consider whether any disclosed information requires incident response
- Review your organization's breach notification requirements
How does this compare to other known traffic analysis attacks?
Whisper Leak targets streamed AI chat replies. Older work targeted other services, such as encrypted video. Standard classifiers work well against streamed AI replies. That is what makes it concerning. An attacker needs a network vantage point and training data. Decryption is not required.