
Whisper Leak: the shape of an encrypted answer was enough to guess the question
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 89 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 97 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-079 tok/s
- OrcaNEWOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 88 tok/s
- OrcaNEWOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 134 tok/s
- OrcaNEWOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 88 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 122 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 69 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 91 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 44 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 194 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 123 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
- OrcaOrca: OrcaCyber Zero 1.52026-10-10$3.00 / $7.50 per 1M tokens · 140 tok/s
- openaiOpenAI: GPT-6.1 Sol2026-09-2955Intelligence
- anthropicAnthropic: Claude Sonnet 5.52026-09-2855Intelligence
- typesafeTypeSafe: Jev 1.132026-09-24$0.04 / $0.00 per 1M tokens · 189 tok/s
On 7 November 2025 Microsoft's Defender Security Research Team published Whisper Leak, a side-channel attack that infers what a person is asking a chatbot from the sizes and timing of the packets carrying the reply — traffic that TLS has already encrypted and that the attacker never decrypts. The technical report went to arXiv two days earlier, on 5 November, under the names Jonathan Bar Or and Geoff McDonald (arXiv:2511.03675). Across 28 commercial language models the researchers report topic classifiers that in most cases scored above 98% AUPRC, and at a simulated 10,000-to-1 class imbalance, 17 of the 28 reached 100% precision at the low recall levels the paper reports. No interception of a real conversation is reported anywhere in the record: the archive files this entry as medium severity with an A confidence grade, no dispute around it, and no recorded real harm. The record itself — its severity grade, source list and confidence assessment — is filed in the AI incident archive, alongside the other network-layer and privacy entries in the same collection.
Why a ciphertext still has a shape
Microsoft's writeup gives the mechanism in the language of cipher design. TLS builds its session on symmetric ciphers, and the two families behave differently at the boundary: a block cipher such as AES in a block mode always emits a multiple of the block size, while a stream cipher such as ChaCha20 or AES-GCM accepts any data length at all. Without compression, the blog states, the ciphertext is the plaintext plus a constant, so the size of each encrypted chunk still tracks the size of what went into it. Language models generate one token at a time and stream those tokens to the client as they are produced, which turns a response into a sequence of network events whose lengths follow the lengths of the tokens and whose spacing follows how fast the model generates them. An observer who can only watch — the blog names a nation-state actor at the internet service provider layer, someone on the local network, or someone sharing the same Wi-Fi router — records that sequence and learns something about the answer without touching the encryption.
The paper is explicit that this is not a break of TLS. The cipher does exactly what it was designed to do; what leaks is metadata that TLS was never built to hide. The authors also place the work in a line of earlier findings: token-length reconstruction by Weiss and colleagues in 2024, timing attacks on speculative decoding by Carlini and Nasr, output-token-count leakage studied by Zhang, Saileshwar and Lie, and cache-sharing timing work by Zheng and colleagues. Whisper Leak's contribution is narrower and more practical than any of those: not reconstructing the answer, but deciding whether the question belonged to one chosen category.

How the experiment was built
The target topic was the legality of money laundering. The researchers generated 100 semantically similar variants of questions on it, held 20 back for testing and trained on the other 80, and mixed them with 11,716 unrelated questions drawn from the Quora Questions Pair dataset. Traffic was captured per model with tcpdump, with positive and negative samples collected in shuffled order and extra spaces inserted between words to avoid prompt caching blurring the results; generation ran at temperature 1.0. Three classifier families were trained on the resulting sequences — LightGBM, a bidirectional LSTM, and a DistilBERT model adapted to treat size and time buckets as tokens — and each was tested on size features alone, timing features alone, and both together.
Performance was scored as area under the precision-recall curve, a choice the paper defends on the grounds that the realistic problem is heavily imbalanced: an observer watching a lot of ordinary traffic for a very small number of interesting conversations should be judged on precision, not on raw accuracy.
The numbers, and what kind of numbers they are
Seventeen of the 28 models cleared 98% AUPRC, several at or above 99.9%, and packet size alone was enough for strong results in most cases. The top of the table is Mistral Large at 100.0%, Microsoft-hosted DeepSeek-R1 at 99.9%, xAI's Grok 3 Mini Beta at 99.9%, and OpenAI's GPT-4o mini at 98.6%. The bottom is not immunity but distance: Microsoft's own Phi-3.5-mini-MoE instruct scored 76.9%, Amazon Nova Lite 71.2%, Amazon Nova Pro 77.5%, the Gemini 1.5 and 2.5 variants between 81.9% and 84.0%, and Anthropic's Claude 3 Haiku 91.2%.

The resistance at the bottom has an explanation the authors offer and do not prove. Batching several tokens into one network event collapses the fine structure a classifier feeds on, and the models that batch hardest did worst in the attack — Gemini 2.5 Pro at roughly 17.7 tokens per event, Gemini 1.5 Flash at roughly 55.3. The paper says it is likely those providers batch for serving efficiency and user-interface refresh rates rather than for side-channel defence, and that batching alone is still insufficient: those same models retained substantial risk at the 10,000-to-1 projection. There is also a gap the paper does not explain. The same model served from two places scored differently — GPT-4o mini through OpenAI at 98.6% and through Microsoft at 93.4%, GPT-4o at 90.1% through Microsoft — which points at the serving stack rather than the weights, but the paper offers no cause.
The 10,000-to-1 figure is the one most likely to be misread. It is a projection from test-set performance onto an imagined stream of 10,000 conversations containing one on the target topic, not a measurement of live traffic. Under it, 17 of the 28 models reach 100% precision somewhere between 5% and 20% recall, meaning that when the classifier raised a flag it was right, at the cost of missing most targets. How much it catches is stated differently in the two primary documents: the arXiv abstract says the authors recover 5–20% of target conversations at 100% precision, while the Microsoft blog says 5–50%. Both are extrapolations, and both documents say so.
What the providers changed
Disclosure began in June 2025, when the researchers notified 28 language model providers, with technical descriptions and proof-of-concept demonstrations on request. By publication, four had shipped something. OpenAI added an obfuscation field to streamed responses — a random-length string of text appended to each response so that token length no longer maps to packet length — and Microsoft Azure mirrored it; the blog states that Microsoft verified Azure's version reduces attack effectiveness "to levels we consider no longer a practical risk". Mistral added a parameter called p with a similar masking effect. xAI is credited with deploying protections without a public description of what it changed. The paper adds the part the blog does not: OpenAI and xAI are recorded as fixing in August 2025, Mistral in September, and Microsoft in October after assessing the risk as moderate, while other providers declined to implement fixes and several did not respond to repeated follow-ups. The paper also notes that all results were collected before any vendor fix shipped, and that publication was held back to November to give willing vendors time to deploy.
The researchers also tested the defences themselves, on a model that was both highly vulnerable and reluctant to be mitigated. Random padding cut AUPRC from 97.5% to 92.9%, batching five tokens per event cut it from 98.2% to 94.8%, and injected noise packets cut it from 98.1% to 93.3% — reductions of 3.5 to 4.8 percentage points, with packet injection costing two to three times the bandwidth. None of the three removes the channel, which is why the deployed fixes are described as mitigation rather than repair.

What is established, and what is not
Established, and checkable by anyone: the paper and its data-collection code are public, the mechanism follows from documented cipher and streaming behaviour, and the four named providers' mitigations exist in their own APIs — an obfuscation field, a parameter named p. Corroboration of the finding itself is thinner than the volume of coverage suggests. The Hacker News, SecurityWeek, Security Affairs and Tech Xplore all reported it between 8 and 11 November 2025, and each account I read in full re-describes the same Microsoft team's experiment from the same preprint rather than reproducing it; the same is true of the further coverage that day from SC Media, Cybernews, TechRadar, CSO, iTnews and CX Today, which I saw but did not read in full. I found no published independent reproduction of the attack, which is a limit on what the coverage adds rather than a contradiction of it.
Not established: that anyone has used this against a real conversation. There is no victim, no organisation, no country and no count anywhere in the primary documents or the coverage reviewed for this article, and the archive record carries no recorded real harm — consistent with the sources rather than a gap in them. Equally not established: how well the attack survives real network conditions, since every figure comes from a controlled capture on one topic; whether multi-turn conversations behave as the authors expect, which they flag as untested; and how the unresponsive providers' traffic behaves today, since nothing published since November 2025 says whether they changed their minds.
One citation detail is worth stating plainly. The archive record's own source is the Japan Information-technology Promotion Agency's AI security bulletin index, a continuously updated Japanese digest that summarises public reporting and that carries a standing disclaimer saying it does not vouch for the accuracy of what it summarises. The December 2025 issue of that bulletin, which I downloaded and read in full, contains no Whisper Leak entry; its eight summary themes are Anthropic's GTG-1002 agent report, indirect prompt injection against agents, attacks on AI-assisted development, MCP vulnerabilities, AI system vulnerabilities, AI reliability failures, careless AI use, and AI for security. The substantive documentation behind this record is therefore Microsoft's 7 November blog post and arXiv:2511.03675, both read in full for this article, and the missing bulletin entry is a fact about the record's citation rather than a doubt about the finding.
What a reader can still do about it
Microsoft's own advice is short and does not pretend to fix the problem: avoid discussing highly sensitive topics with a chatbot while on an untrusted network, use a VPN for a layer of protection in transit, prefer providers that have implemented mitigations, and use non-streaming modes where a provider offers one, since streaming is the mechanism. The structural point underneath it is that exposure here is a property of the provider's serving stack, not of the user's settings — two services running the same weights can sit at opposite ends of the table above depending on how they batch and what they append to a stream.
For providers, the researchers' own conclusion is the uncomfortable one: every defence they tested trades some streaming performance or bandwidth for a partial reduction in a channel that exists because responses are generated and delivered one token at a time. Whisper Leak does not have a patch in the sense a memory-safety bug has one. It has a set of serving decisions, and a measurement of how much each of them buys.
