September 16, 2026 — Aida Baradari's company Deveillance released Kalypta, described as "the first app to block AI notetakers in your meetings." The pitch, in Baradari's own words: "With Kalypta, you become inaudible to AI. Your call continues normally." It's the software counterpart to the company's earlier hardware product, Spectre I, which physically stops microphones from picking up sound.
TL;DR — what people are asking
| Question | Answer |
|---|---|
| What does Kalypta do? | Reshapes your mic audio in real time to defeat AI transcription |
| Does it affect the human listener? | No, per Deveillance — speech "stays clear for humans" |
| What AI notetakers does it target? | Named: Granola, Wisprflow, Cluely; trained against Whisper, NVIDIA Canary, "and more" |
| How effective is it? | Deveillance claims ~2 of 3 words fail to transcribe accurately, on average |
| Where does it run? | A small model, locally, on your own device |
| Is there a hardware version? | Yes — Spectre I, a physical microphone-blocking device, came first |
| Is the effectiveness claim independently verified? | No — it's Deveillance's own reported benchmark |
| What's the underlying concern it's addressing? | "Ambient" AI surveillance — recording/transcription happening without full visibility or consent |
The problem Kalypta is built to solve
AI meeting notetakers — Granola, Otter-style tools, and increasingly agentic browser assistants like Cluely — have become common enough in calls that "is this being transcribed?" is now a live question in a lot of meetings, not a paranoid one. Deveillance's framing leans into that directly: "We live in the age of ambient technology and we have lost sight & agency over where AI is active in our lives."
That's a real, growing category of concern, distinct from but related to the browser-agent autonomy questions explainx.ai has covered around Claude in Chrome and shadow AI use inside workplaces — employees or meeting participants running AI tools without everyone in the room knowing, or being fully aware of what's captured and where it goes afterward.
How it claims to work
Deveillance's technical explanation, from the launch thread: Kalypta "runs a small local model on your device, reshaping audio in real time. It is trained against the same transcription architectures as Whisper, NVIDIA Canary, and more." That's an adversarial-audio approach — rather than encrypting or blocking sound outright, the model perturbs the audio signal in ways specifically tuned to confuse the acoustic models behind automatic speech recognition (ASR), while staying intelligible to the human auditory system, which processes speech very differently from a neural ASR pipeline.
This is conceptually similar to adversarial-example techniques used against image classifiers — small, targeted perturbations invisible or near-invisible to a human that reliably break a specific model family — applied to audio instead of pixels. The claimed generalization ("the same transcription architectures as Whisper, NVIDIA Canary, and more") suggests the perturbation targets shared characteristics of transformer-based ASR encoders rather than one specific vendor's model, though Deveillance hasn't published the technical paper behind the claim.
The demo and the reported numbers
Deveillance posted a demo from Deveillance, described as showing "a conversation between two people... watch the live transcription fall apart" in real time. Alongside the demo, the company reported:
- ~2 of 3 words fail to transcribe accurately, on average, across its internal benchmarks.
- The model "is improving weekly," implying the current effectiveness figure is a snapshot, not a stable number.
- Testing was done against "established intelligibility metrics" for the human-clarity side, plus Deveillance's own internal benchmarks for the AI-defeat side — neither is named specifically in the launch thread.
None of these figures have been independently reproduced by a third party as of this writing. Treat them as a vendor claim worth testing yourself before relying on Kalypta for anything where the stakes of a failed block are meaningful — legal, medical, or confidential business conversations.
From hardware to software: Spectre I and the Deveillance thesis
Kalypta isn't Deveillance's first product. The company's earlier device, Spectre I, takes the more blunt-instrument approach: physically preventing a microphone from picking up sound at all, rather than adversarially confusing a transcription model after the fact. Positioning Kalypta as software built "together with" Spectre I signals Deveillance's actual thesis — this is a company building a portfolio of tools around giving people control over when and how they're audible to AI systems, not a single-app play.
That's a notably different angle from most AI-privacy tooling explainx.ai has covered, which tends to focus on protecting data after capture — blurring faces or license plates in video, or auditing what an AI agent already recorded. Kalypta and Spectre I both intervene before capture, which is a meaningfully different privacy posture: nothing usable ever reaches the AI system in the first place, rather than being captured and then redacted or deleted.
The adversarial-audio arms race this implies
If Kalypta's approach genuinely works as described, it sets up a dynamic familiar from adversarial machine learning generally: a defense tuned against a specific model family degrades once that model family updates, and an attacker (or in this case, defender) has to keep retraining against the moving target. Deveillance's own "improving weekly" framing is itself an acknowledgment of this — the company is signaling ongoing investment in staying ahead of whatever Whisper, Canary, or other ASR architectures do next, rather than shipping a one-time fix.
This mirrors a pattern that's played out repeatedly in adversarial-example research against image classifiers: a perturbation crafted against one model generation often transfers to related models but degrades against genuinely new architectures, especially ones trained with adversarial examples specifically in mind. If AI notetaker vendors start training their ASR models to be robust against exactly the kind of audio reshaping Kalypta performs — a realistic response once a product like this gains real adoption — the current effectiveness numbers could shift meaningfully within months, in either direction. Neither side of that race has an obvious permanent advantage, which is worth keeping in mind before treating Kalypta as a durable, install-once solution rather than an ongoing arms race its maker will need to keep winning.
Where this fits in the broader AI-privacy product category
Kalypta arrives at a moment when "does this AI system know I don't want it capturing me" has become a genuine product category rather than a niche concern. It sits alongside a growing list of tools aimed at giving individuals agency over ambient AI capture rather than relying on policy or platform-level opt-outs — physical camera covers, audio-jamming hardware like Deveillance's own Spectre I, and increasingly, software-level countermeasures targeting the specific AI models doing the capturing. What makes Kalypta notable within that category is the specificity of its target: rather than a general privacy tool, it's built against a named, common class of workplace software (meeting notetakers) that most people encounter routinely, often without a clear signal of whether it's active in a given call.
Honest limitations
- Vendor-reported effectiveness only. The "2 of 3 words" figure and the "established intelligibility metrics" claim haven't been reproduced independently.
- "Improving weekly" cuts both ways. It implies active development, but also that today's launch-day performance is explicitly not the final state — a model that degrades ASR accuracy today may be defeated by an ASR model update tomorrow, and vice versa.
- No word on pricing, platform availability, or whether it works across desktop and mobile calling apps uniformly.
- A live arms race is implied but not addressed. If AI notetaker vendors adapt their models to Kalypta's perturbation patterns (the way adversarial-defense research usually escalates), Deveillance would need to keep updating Kalypta to stay ahead — a maintenance burden with no guaranteed end state.
- A social/consent question, not just a technical one. Using audio-reshaping software in a meeting where other participants are relying on an AI notetaker for legitimate reasons — accessibility, non-native-language support, accurate minutes — raises a disclosure question independent of whether the tool works as claimed.
What this means for what you build or pay
If you build or deploy AI meeting notetakers: Kalypta is a live signal that adversarial-robustness against real-world audio perturbation is now a competitive requirement, not a hypothetical one — the same way prompt-injection resistance became a requirement for agentic tools after enough public demonstrations.
If you're a knowledge worker concerned about ambient AI capture: Kalypta and Spectre I are worth testing in a low-stakes meeting before trusting them in a sensitive one, given the unverified effectiveness claims and the "improving weekly" caveat.
If you're evaluating AI privacy tooling broadly: compare pre-capture tools like Kalypta against post-capture redaction tools like video face-blurring — they solve different halves of the same underlying problem, and a serious privacy posture likely needs both, not one or the other.
Related on explainx.ai
- Shadow AI in the workplace: the privacy risk guide
- Claude in Chrome: features, access, and safety
- How to blur faces in video for AI privacy protection
- How to blur license plates in video with AI
- EU driver-facing camera law: a privacy guide
- What is AI ethics? A complete guide
Feature claims and effectiveness figures reflect Deveillance's own September 16, 2026 launch announcement on X — these are vendor-reported figures, not independently verified benchmarks, as of publication.