Do you care about AI privacy? If so, you'll love our latest open-source release from our private AI stack! Introducing twoway - an encrypted request-response messaging library built on HPKE. Repo and technical breakdown links in the comments

Policy’s not gonna save us - Meta’s American Technology Excellence Project will kill any AI privacy legislation they see

Most AI privacy content is noise. The Dead Drop is signal - sharp takes and deep dives from our team for you. First 10 newsletter signups get exclusive privacy swag (ever wanted to block facial recognition?) What AI/privacy topic should we hit?

Creepers, cheaters, and privacy besiegers, you’re done! Don’t Record Me will be ready soon, we let you choose when AI transcribers can capture your conversation. Big thanks to @sfstandard.com for the shoutout! Sign-up link here: dontrecord.me

today we're bumping Don't Tap The Glass by Tyler the Creator. the impact of sousveillance on club culture, and culture altogether, is degenerative -- one of the many reasons we care about privacy. privacy is upstream of trust, which is upstream of joy

🛑 People feed sensitive info into AI models every day. Where does it go? Who sees it? Can it be deleted? Without clear answers, your privacy’s at risk. CONFSEC guarantees no visibility, no retention—provably. Watch our CEO talk about #AIPrivacy on TBPN, link on the comments. #Privacy #LLMSecurity

We're out of stealth! Our CEO will be live today to talk AI Privacy on TBPN at 1:45PM PT🎙️ 🔗 www.youtube.com/@TBPNLive Check out our announcement on TechCrunch 😎 📰 techcrunch.com/2025/07/17/c... #AIPrivacy #EnterpriseAI #LLMSecurity #DataProtection

Confident Security, ‘the Signal for AI,’ comes out of stealth with $4.2M

This is wild 😳: Meta just fixed a bug that could leak your AI prompts and responses. Turns out our chats weren’t as private as we thought. This is why zero data retention matters. #AI #Meta #Privacy #PromptLeak #LLMSecurity

We loved Helen Nissenbaum's discussion on @techpolicypress.bsky.social's ‬podcast: 🔸Privacy isn’t about hiding—it’s about the appropriate flow of information. 🔸Obfuscation is a response to broken systems, not bad behavior. 🔸When regulation fails, resistance becomes a right. #Integrity #AIPrivacy

𝗡𝗲𝘄 𝗬𝗼𝗿𝗸 𝗷𝘂𝘀𝘁 𝗽𝗮𝘀𝘀𝗲𝗱 𝘁𝗵𝗲 𝗥𝗔𝗜𝗦𝗘 𝗔𝗰𝘁: – 📊 Risk assessments for AI decisions – 🛡️ Transparency + civil rights protections A win for responsible AI implementation. #RAISEAct #InferencePrivacy

Is a privacy-friendly browser really too much to ask?

@meredithmeredith.bsky.social highlights the real risks of agentic AI: lack of oversight, accountability, and security. ⚠️ Smart agents aren’t necessarily a threat. Unchecked inference is. 🔐 We need private-by-design AI infra. #AgenticAI #InferencePrivacy

The AI privacy gap no one talks about? Inference data. 🧠 You locked down your training set. Great. But what about live prompts? 💬 Support tickets with PII 📄 Legal drafts 🏥 Medical data 💰 Financial details Most systems log it all like it’s nothing. It’s not. It’s everything. #AI #InferenceSecurity

🧠 Imagine this: An employee pastes customer data into their favorite AI tool—no review, no approval. That data is now sitting on a third-party server you don’t control. When AI tools operate under the radar, who’s keeping your data safe? 𝗧𝗵𝗶𝘀 𝗶𝘀 𝗦𝗵𝗮𝗱𝗼𝘄 𝗔𝗜: 𝘂𝗻𝘃𝗲𝘁𝘁𝗲𝗱, 𝘂𝗻𝗰𝗵𝗲𝗰𝗸𝗲𝗱, 𝗻𝗼 𝗴𝘂𝗮𝗿𝗱𝗿𝗮𝗶𝗹𝘀.

They can't keep getting away with this😩

📊 𝗠𝗼𝘀𝘁 𝗔𝗜 𝘃𝗲𝗻𝗱𝗼𝗿𝘀 𝗹𝗼𝗴 𝗺𝗲𝘁𝗮𝗱𝗮𝘁𝗮—𝘄𝗵𝗼 𝘆𝗼𝘂 𝗮𝗿𝗲, 𝘄𝗵𝗮𝘁 𝘆𝗼𝘂 𝘀𝗲𝗻𝘁, 𝗮𝗻𝗱 𝘄𝗵𝗲𝗻 𝘆𝗼𝘂 𝘀𝗲𝗻𝘁 𝗶𝘁. Only CONFSEC guarantees that prompts and metadata aren’t stored or visible.

Trust in “AI-powered” tools is fading—especially when users don’t know where their data ends up. For sectors like finance, healthcare, and law, that’s not just a UX problem. It’s a compliance one. Privacy can’t be a checkbox. It has to be verifiable.

🧠 As AI adoption accelerates, Varonis has looked at how prepared organizations really are for the risks they might bring. They examined 1,000 companies across [healthcare, etc.] and found ghost accounts, weak MFA, and sensitive files exposed. These aren’t new problems—AI just exposes them further.

𝗤𝘂𝗲𝘀𝘁𝗶𝗼𝗻 𝗳𝗼𝗿 𝗔𝗜 𝗯𝘂𝗶𝗹𝗱𝗲𝗿𝘀: 𝘋𝘰𝘦𝘴 𝘴𝘵𝘰𝘳𝘪𝘯𝘨 𝘦𝘷𝘦𝘳𝘺 𝘱𝘳𝘰𝘮𝘱𝘵 & 𝘳𝘦𝘴𝘱𝘰𝘯𝘴𝘦 𝘳𝘦𝘢𝘭𝘭𝘺 𝘩𝘦𝘭𝘱 𝘺𝘰𝘶 𝘤𝘳𝘦𝘢𝘵𝘦 𝘵𝘩𝘦 𝘣𝘦𝘴𝘵 𝘱𝘦𝘳𝘧𝘰𝘳𝘮𝘪𝘯𝘨 𝘮𝘰𝘥𝘦𝘭? 🤔 Follow the thread below ⬇️

🔍 𝗣𝗿𝗶𝘃𝗮𝘁𝗲-𝗔𝗜 𝗗𝗲𝗽𝗹𝗼𝘆𝗺𝗲𝗻𝘁 𝗖𝗵𝗲𝗰𝗸𝗹𝗶𝘀𝘁 Ship trust, not risk. 🧵👇

📊 𝗖𝗵𝗮𝘁𝗚𝗣𝗧, 𝗖𝗹𝗮𝘂𝗱𝗲, 𝗚𝗲𝗺𝗶𝗻𝗶, 𝗠𝗲𝘁𝗮 𝗔𝗜, 𝗣𝗲𝗿𝗽𝗹𝗲𝘅𝗶𝘁𝘆 & 𝗠𝗶𝘀𝘁𝗿𝗮𝗹—train on your prompts by default. For regulated data, that’s a deal-breaker. ⚙️ CONFSEC wraps your model: no visibility, no retention, provable privacy. 🔒 Link to our scorecard in the first reply.

🚨 𝗔𝗜 𝗽𝗿𝗶𝘃𝗮𝗰𝘆 𝗳𝗮𝗶𝗹𝘀 𝗳𝗮𝗹𝗹 𝗶𝗻𝘁𝗼 𝟰 𝗯𝘂𝗰𝗸𝗲𝘁𝘀. 1️⃣ Inference leaks 2️⃣ Training grabs 3️⃣ Scope creep 4️⃣ Trace spills If your guardrails miss even one, data walks. 𝗖𝗢𝗡𝗙𝗦𝗘𝗖 = no visibility, no retention, provable privacy.

🧵 𝗜𝗻𝗰𝗼𝗴𝗻𝗶’𝘀 𝟮𝟬𝟮𝟱 𝗽𝗿𝗶𝘃𝗮𝗰𝘆 𝗿𝗮𝗻𝗸𝗶𝗻𝗴 𝗲𝘃𝗮𝗹𝘂𝗮𝘁𝗲𝗱 𝟵 𝗚𝗲𝗻𝗔𝗜 𝗽𝗹𝗮𝘁𝗳𝗼𝗿𝗺𝘀 𝗼𝗻 𝘁𝗿𝗮𝗶𝗻𝗶𝗻𝗴, 𝘁𝗿𝗮𝗻𝘀𝗽𝗮𝗿𝗲𝗻𝗰𝘆, 𝗮𝗻𝗱 𝗱𝗮𝘁𝗮-𝘀𝗵𝗮𝗿𝗶𝗻𝗴 𝗽𝗿𝗮𝗰𝘁𝗶𝗰𝗲𝘀. Follow the thread below!

🧵 “We regex the PII.” 🧵 “Our vendor promised not to log anything.” That’s not privacy. That’s ✨faith-based security✨. Regex fails silently. Vendors change policies. Inference data will leak if your architecture relies on trust alone. 𝗖𝗢𝗡𝗙𝗦𝗘𝗖 solves it: no logs, no retention, provable privacy. 🔐

🍏 PCC showed what private inference can look like. 🔐 CONFSEC makes it real—on any cloud, with any model, for everyone. 🧩 Drop-in wrapper. Same API. No rebuilds. 📜 Zero logs. Zero retention. And we can prove it. 𝗖𝗢𝗡𝗙𝗦𝗘𝗖 = Enterprise-grade privacy for AI. #PrivateAI #PCC #AIprivacy #DecentralizedAI

“𝗝𝘂𝘀𝘁 𝘂𝗻𝗹𝗲𝗮𝗿𝗻 𝗶𝘁 𝗹𝗮𝘁𝗲𝗿” = ⏰💸 ➡️ Harvard/MIT: On 128 prod models, 91% needed retraining due to data drift. ➡️ UNC/DeepMind: Exact-unlearning request triggers multiple new fine-tune passes. 🏁 Go private from day one. CONFSEC enclaves keep raw data out of the weights—no treadmill later. #Privacy #AI

🤖⚠️ Anthropic’s “goal-conflict” test: a fake competitor asks for trade secrets, model obeys. Some engines leaked data on every single run. 😬 Link in the first reply. #LLMSecurity #AI

🧵 𝗖𝗢𝗡𝗙𝗦𝗘𝗖 𝗙𝗔𝗤 𝗧𝗵𝗿𝗲𝗮𝗱 𝘘: What makes 𝗖𝗢𝗡𝗙𝗦𝗘𝗖 different from a “secure cloud”? 𝘈: Your LLM runs inside a confidential-computing enclave. Remote attestation proves neither we—nor the cloud host—can read prompts or outputs. 👇

WhatsApp’s AI helper accidentally shared phone numbers. 🤦‍♂️ Useful—until it spills private info. Rule #1: limit what your model can access. 🔒 𝗖𝗢𝗡𝗙𝗦𝗘𝗖 does exactly that. #AI #Privacy

🔍 Three privacy levers every LLM team should know: filter, redact, enclave. Together you can control who sees data, when, and what they see. Follow the thread below ⬇️ #AI #DataPrivacy #LLMSecurity

📰 In Jason Meller’s latest Hacker News piece, GitGuardian reports 24 million hard-coded secrets on public GitHub last year + repos using Copilot leaked roughly 40% more credentials. Best practices: short-lived tokens, secret scanners, no keys in chat. #DevSecOps #SecretsManagement

🚨 𝗕𝗲𝘄𝗮𝗿𝗲 𝗘𝗰𝗵𝗼𝗟𝗲𝗮𝗸: one crafted email let attackers pull data from Microsoft Copilot—zero clicks, pure prompt abuse. If an email can hijack your agent, guardrails aren’t optional. #PromptInjection #LLMSecurity #AI

🤔 Prompt injection got you nervous? ✍️🛡️ A fresh paper lays out six reproducible patterns that fence in agent actions and curb risk. Worth a look if you’re building with LLMs. 📉🤖 #PromptInjection #AI #Security

🤖 Building an AI product? Cool. 🔒 Guaranteeing your users’ data is unreadable—even to you? Cooler. 𝗨𝘀𝗲 𝗖𝗢𝗡𝗙𝗦𝗘𝗖: enterprise-grade, audit-backed wrapper for any inference engine. Deploy on-prem, in any cloud, or both. Privacy receipts included. #AI #DataPrivacy #ConfidentialComputing #ZeroLogging

🍏 Apple’s new Foundation Models Framework is a quiet game-changer. 📱 On-device 🧠 No cloud training 🕵️ No scraping AI without surveillance. Usability and privacy? More of this, please. #OnDeviceAI #Apple #Privacy #FoundationModels

🤖 AI-powered apps are everywhere — but some are asking for way too much. 🗂️ File access 📇 User directory 📆 Calendar permissions 𝗖𝗼𝗻𝗳𝗶𝗱𝗲𝗻𝘁 𝗦𝗲𝗰𝘂𝗿𝗶𝘁𝘆 keeps them in check, making sure their models can use your data, but never keep it. 🔐 #AI #Privacy #CyberSecurity #SaaS #ZeroTrust

Apple is opening up access to its offline AI models. 🤖 ✅ Private ✅ Fast 🚨 But also running inside your apps This is a huge shift in where AI runs — and who controls what it touches. Link in the replies below ⏬ #AI #Privacy #OnDevice #Apple

🧠 Zoom, Slack, that new scheduling tool — they all ask for access. You click ‘Allow’. Suddenly, one can modify your Google Drive. No hacks. Just over-permissioned apps. 𝗖𝗼𝗻𝗳𝗶𝗱𝗲𝗻𝘁 𝗦𝗲𝗰𝘂𝗿𝗶𝘁𝘆 𝗵𝗲𝗹𝗽𝘀 𝘆𝗼𝘂 𝘀𝗮𝘆 𝘆𝗲𝘀 𝘁𝗼 𝘁𝗵𝗲 𝗿𝗶𝗴𝗵𝘁 𝘁𝗵𝗶𝗻𝗴𝘀. 🔐 #CyberSecurity #OAuthSecurity #ZeroTrust #IAM #DataPrivacy