Hey @fouadmatin @thsottiaux, I am a security and privacy researcher. At @Brave I work on critical products like Brave Wallet - where a missed vulnerability can mean users lose funds. My personal projects, like sitebehavior.org, are fully open source. Why does Daybreak need a government ID so I can analyze code I already maintain in public? I get the dual-use concern. Attackers aren't filling out KYC forms. They are using open-weight models. The friction falls on defenders, the people you say you want to help. Does that seem fair to you?

Alongside the exploit discussion, there is a longstanding privacy issue for people who never touch the car. With Dashcam on, a Tesla records its surroundings while driving. With Sentry Mode on, it watches while parked and can save footage when it detects a security event. Tesla says Dashcam and Sentry recordings are saved in the car or on a USB drive, never on its servers. And, it collects operational and diagnostic data, receives camera clips owners opt into sharing for fleet learning, and says it can receive separate camera recordings following serious safety-critical events. The owner chose to turn on recording. The people walking past did not. What protects them when that footage is kept, shared, or handed to police?

Perplexity sent the full text of my test searches to Datadog when the search arrived in the page URL. I captured the uploads live. [That is how browsers send searches when Perplexity is the search engine.] I usually break down privacy policies. This time, I followed what actually left the browser. I used harmless synthetic queries against the live website, without changing Perplexity’s code or stubbing responses. The query was stripped from the standard page-URL field sent to Datadog, but remained in two nested performance fields. After the initial captures, I ran the follow-up tests shown in the image. It includes the conditions, results, and limits - including the runs where I did not observe query transmission. One upload happened with Global Privacy Control enabled. Perplexity describes GPC and its cookie choices as opt-outs from targeted advertising. What happened to the data after Datadog accepted it remains unknown. @perplexity_ai @datadoghq @AravSrinivas @kpolley

DeMicrosoft 101: practical replacements for the Microsoft ecosystem Outlook → @ProtonMail, @TutaPrivacy Microsoft Photos → @enteio Photos Bing → @brave Search Edge → @brave Microsoft Teams → @signalapp, @SimpleXChat OneNote → @notesnook, @StandardNotes OneDrive → @enteio Locker, @ProtonDrive Microsoft Password Manager → @Bitwarden, @Proton_Pass Microsoft Authenticator → @enteio Auth, @Bitwarden Auth, @ProtonPrivacy Authenticator Outlook Calendar → @ProtonPrivacy Calendar, @TutaPrivacy Calendar Outlook Contacts → @ProtonPrivacy Contacts, @TutaPrivacy Contacts Microsoft Copilot → @brave Leo, @asklumo Microsoft Translator → @brave Translate Microsoft Teams (video) → @brave Talk, @ProtonMeet Windows → secureblue Microsoft 365 → @fileverse, @LibreOffice, @cryptpad SwiftKey → FlorisBoard, Apple Keyboard

Google Street View can show ANYONE your front door / driveway / visible security cameras without them ever visiting your street. You can request a blur but that means giving Google proof of address and an email linked to your claim that you own or rent the home. If you go ahead, minimize what you send. Use a forwarding email alias (like @SimpleLogin) you use nowhere else. This keeps your primary email out of the form [it does not make the request anonymous]. Use a utility bill dated within 6 months. Redact any sensitive information. Leave the address and document date visible. Do NOT upload a government ID. Google says it permanently deletes uploaded documents after 60 days. Google’s requirements: support.google.com/maps/answ… To request it: Google Maps → Street View → “Report a problem” → “My home.” Upload the redacted proof, enter the address and submit. Once applied, the blur is permanent. Your address remains searchable.

126 million people use @Brave every month. 53 million use it every day. It started in 2015, when @BrendanEich and @brianbondy set out to build a browser that works for YOU - NOT for whoever is tracking you. Two minutes, 2015 to today. Easter egg: the main theme of the score is the word BRAVE encoded into musical notes.

Hey @shl, I am a self-proclaimed DOGE for privacy. I like helping the American people protect their data and stay safer online. I looked at the IRS app's public privacy disclosures. Here's what I found. On Google Play, the developer-provided label lists crash logs, app interactions, and device or other IDs as data the app may collect and may share with other companies or organizations. For device or other IDs, it lists Analytics as the purpose. The App Store label says usage and diagnostic data may be collected for analytics but are not linked to your identity. It separately lists a user ID linked to identity for app functionality. Both listings link to the general IRS privacy policy. It covers IRS.gov and its online applications, but does NOT explain the mobile app's analytics practices. The IRS's public PIA index links “The IRS app” to an IRS2Go assessment approved in June 2020. That document predates this release and does NOT discuss analytics. With the app now offering tax records, notices, and bank account updates, could the IRS prominently link the current mobile app privacy notice and applicable assessments from its app page and store listings? It would help to know which providers receive which data, which identifiers are used and whether they can be linked to taxpayer accounts, how long the data is kept, and what users can turn off. It would also help to confirm that tax identifiers, bank details, document contents, and authentication tokens are excluded from analytics and crash reports. These are requests for clarity and verification. I only reviewed public pages and their expanded disclosures. The screenshots highlight the relevant disclosures.

3 copies of your data can still be vulnerable to the same disaster. That's what stood out to me about the AWS data loss in Bahrain and the UAE. AWS said it couldn't restore access to data hosted exclusively in its Bahrain region or in an affected UAE availability zone. 3-2-1 backup rule is a good place to start: 3 copies, on 3 different types of media, with 1 copy offsite. For me, the real question is what could take out every copy at once. Multiple availability zones give you protection within a cloud region. Surviving the loss of that region requires a recovery plan that reaches beyond it. Even then, the location matters. If you need your data to stay in the US, you have more room to work with. AWS has 4 commercial US regions: N. Virginia, Ohio, N. California, and Oregon. Keeping data in the country doesn't force you into one region. Pair an eastern region with a western one. Ohio and Oregon, for example, are far apart and on different major power grids, which lowers the odds of one regional disaster hitting both. For data you can't afford to lose, I would also keep a recoverable copy outside AWS, still in the US. A different provider's US storage or an encrypted offline copy you control adds protection against a problem affecting your provider. Make sure you can access that copy and its encryption keys independently.

Before buying a new router, check whether yours still receives security updates. If it does and the WiFi works well throughout your home, I would keep it. If you need a replacement, here are two options, depending on how much setup you are prepared to do. 1. For the simpler setup: GL.iNet Flint 2 (GL-MT6000). This would be my pick between the two for a beginner. It has Wi-Fi 6 and can be set up locally through a browser without creating a cloud account. Start with GL.iNet’s supplied firmware. Set a strong administrator password, leave cloud remote management disabled unless you specifically need it, and install stable updates through System -> Upgrade. Leave preview releases disabled. You are relying on GL.iNet for that firmware and its updates. It is based on OpenWrt, but it is not the same firmware maintained and released by the OpenWrt project. 2. For someone willing to learn network configuration: OpenWrt One. It comes with OpenWrt, an opensource OS for routers [already installed]. I would choose it for someone who specifically wants that software and is prepared for more hands-on setup. It has one 2.5GbE WAN port for the internet connection and one 1GbE LAN port for wired devices. To connect multiple wired devices, add a network switch to the budget. For most people, I would prioritize a router they can configure and keep updated over one that becomes a networking project just to get online.

Researchers got two already-compromised computers to communicate through heat, without a network connection between them. One computer varied its heat output. The other detected those changes using built-in temperature sensors. In the 2015 BitWhisper experiment: - The machines were up to 40 cm apart. - Transfer rates were 1–8 bits per hour. - Both computers were already compromised. @polynoamial cites this academic example when discussing AI agents that are supposed to be isolated and independent. Coordination can require very few bits, so even a slow channel could matter. He calls air-gapping an extremely strong safeguard. His argument is that absolute guarantees about isolation are difficult, which is why safety protocols need independent layers of defense. One lesson from the @huggingface incident was that too much trust had been placed in sandbox isolation, with too few independent safeguards. What could those defenses look like? For physical communication channels, researchers discuss physical separation, zoning and activity detection. Their effectiveness depends on the channel and the environment being protected. Related research: “AI Control” paper tests monitoring, auditing and trusted-model editing against deliberate backdoor insertion in a limited coding setting. The broader argument is that safeguards can buy time while the alignment problem still needs solving. The question I would like to understand better: if isolation fails, which other safeguards still hold, and how have they been tested against agents trying to defeat them?

Agreed @BorisMPower. AI conversations deserve strong legal protection. OpenAI can build privacy into its products by design. Start with the defaults users live with right now. 1. Training is enabled by default on personal Free, Plus and Pro workspaces. Paying for Plus or Pro does not automatically exclude conversations from training. 2. The opt-out has a feedback exception. OpenAI's documentation says that even after opting out, submitting thumbs-up or thumbs-down feedback can make the entire associated conversation eligible for training. 3. For eligible Free and Go users receiving ads - turning off ad personalization still allows the current conversation to inform ad selection. Using conversation content to pick an ad is still a choice OpenAI makes. I prefer privacy by design over privacy by policy.

Muse's "secure VM" does NOT mean Meta cannot access your data. Training on your conversations is also enabled by default. According to Meta’s own security write-up: - Meta retains the ability to access VM data when necessary to support, secure or operate the service. Access by Meta personnel is restricted through "operational" policies. - Conversations, tool calls and subagent handoffs are used for model training after removing key personally identifiable information. Meta calls this training arrangement "a good default."

I do NOT consider stock Firefox privacy-respecting by default. I reviewed Mozilla’s privacy notice and technical documentation. These are the details users should know. - Search suggestions are enabled by default in normal browsing. They send your search text to your selected search provider as you type, before you submit it. Where Google is the default, those requests go to Google. - Firefox also collects telemetry by default. Mozilla documents client and profile-group identifiers in some telemetry pings, explicitly designed to correlate reports. Those identifiers allow telemetry reports to be linked to a browser profile. That alone does not establish that browsing history is being uploaded, but it does make linkability a concrete privacy concern. - Turning off the main technical and interaction data setting does NOT disable the separate daily usage ping. That ping has its own opt-out. Mozilla says it excludes browsing history. Users should not have to discover this distinction after believing they have disabled usage reporting. - There is also a difference between Firefox’s default tracking protection and its stricter configuration. In Standard mode, “tracking content” blocking applies only to Private Windows. Extending that protection to ordinary windows requires Strict or Custom mode. Firefox has meaningful protections, including Total Cookie Protection enabled by default. Those deserve credit. My criticism is specific: search text leaves the browser while you type, some telemetry supports correlation across reports, usage reporting has separate controls, and broader tracking-content blocking requires a settings change. These defaults fall short of what I expect from a browser marketed as private by default. Sources from Mozilla in the reply.

DeGoogle 101: practical replacements for the Google ecosystem Gmail → @ProtonMail Google Photos → @enteio Photos Google Search → @brave Search Chrome → @brave Google Messages → @signalapp Google Keep → @notesnook, @StandardNotes Google Drive → @ProtonDrive, @enteio Locker Google Password Manager → @Bitwarden, @Proton_Pass Google Authenticator → @enteio Auth Google Calendar → @ProtonPrivacy Calendar Google Contacts → @ProtonPrivacy Contacts Play Store → @accrescentapp + Obtainium Google Gemini → @brave Leo, @ProtonPrivacy Lumo Google Maps → @osmandapp + @OrganicMapsApp Google Translate → @brave Translate Google Meet → @brave Talk, @ProtonPrivacy Meet Google Pay → @Zcash Android → @GrapheneOS Google Workspace → @LibreOffice, @fileverse, @cryptpad

I built Site Behavior Lab to turn a website visit into evidence you can inspect. v0.6 update brings: - Clearer reports that link findings to evidence - Better comparisons, site history, search and exports - Clearer coverage limits and uncertainty - Faster CI and deployments, with verification intact It’s open source and runs real browser scans on Cloudflare. Hosting it reliably is the part I can’t afford to keep funding on my own. @Cloudflare @github, are there OSS credits, sponsorship programs or teams you’d point me to? I’d also appreciate advice from other maintainers on grants or potential sponsors. I’m open to investment if there’s a fit. If you know someone who supports opensource privacy tools, an introduction or a share would help.

Take app, browser, and OS updates more seriously than ever. The economics of exploitation are changing fast. Brave desktop build 1.94.121 includes the Chromium fix for CVE-2026-85046 [a V8 vulnerability for which Google confirms an exploit exists in the wild]. Now connect that with OpenAI’s latest results. GPT-6 Astra scored 100% on ExploitBench. This benchmark gives an AI agent a known V8 vulnerability, source code, debugging tools and the patch (but no reference PoC). It measures progress toward arbitrary code execution. V8 is the engine involved in the browser bug above. My takeaway: as exploit development gets easier, delaying security updates becomes harder to justify. A published patch gives defenders a fix. It can also give attackers clues about how to exploit devices that haven’t updated. If Brave says “Relaunch to update,” do it now.