If you've got Private Relay baked into your OpSec setup, worth knowing: researchers found a set of WebKit bugs (the passkey/WebAuthn flow, DNS prefetching, and WebTransport) that let a site quietly grab your real IP address even while Private Relay is on. No prompt, no indication it happened. It also hits OnionBrowser on iOS.

The practical bit: Private Relay only ever covered Safari traffic, never your whole device, and this shows even that promise doesn't hold in every case. Apple's confirmed it's investigating but hasn't given a fix date. If you're relying on it as your only layer, this is a good moment to pair it with something that tunnels all traffic instead of just Safari's.

Second Apple privacy feature to fail this way in as many months (Hide My Email got the same treatment from 404 Media last month).

Someone spent thirteen hundred messages asking ChatGPT to check his theory about black holes. The bot told him he was ahead of 90% of published physics. He wasn't. It never said so.

Henk van Ess ran the numbers on the AI chatbot conversations that got indexed publicly (the story behind how those got exposed is worth its own read, and it's the piece I flagged last time as a bit stale). This is the follow-up: 112,000 conversations, 1.8 million messages, across ChatGPT, Claude, Copilot, and Gemini, all aimed at one question — does the model pull people deeper into a delusion, or do people do that themselves? He's got data on whether it's one company's problem or all of them, and whether you can actually spot it happening to you in real time.

Part of it's behind Digital Digging's paywall, but the free portion covers the core findings. Good one for anyone thinking about how AI chat logs get used (or misused) as an evidence source, or just watching what these tools do to people's epistemics.

OCCRP and the Straits Times just did something satisfying: they took a US federal indictment that anonymized seven Prince Group co-conspirators, and they unmasked one of them anyway. "Co-Conspirator-2" is allegedly the group's chief enforcer, the guy who bribed cops and officials to keep the Cambodian scam empire's compounds running. The indictment kept his name out, but it couldn't scrub the personal details.

The identification hinges on something almost mundane: the indictment says Co-Conspirator-2 chaired a specific subsidiary between 2017 and 2022. Cambodian corporate records show exactly one person held that seat during that window — Chen Sokly. From there it's a straight line through California property records, a Mauritius company registry, a Singapore superyacht dealership, and even a 2018 party video showing him mingling under an alias. None of it is exotic. It's just patient cross-referencing of public filings against a document that tried hard not to name him.

Worth reading if you ever have to work from a redacted or anonymized source document — this is a clean playbook for how much a timeline detail can give away.

Indicator pulled every tool listed across 19 public OSINT toolkits — more than 12,000 entries — and ranked them by how often they actually show up, rather than trusting any single curator's taste. 57 tools made the cutoff (ties all included rather than broken arbitrarily), and notably, every one of them is free or has a working free tier.

The more useful part for daily use is what they built to generate the list: OSINT Navigator, a free app that aggregates all 19 toolkits into one searchable interface. Type something like "how do I find the owner of a website" and it surfaces the relevant tools, which toolkits recommend them, and a summary of your options. It's rebuilt weekly against a public dataset, so it doesn't go stale the way a static list does.

Monitori — an OCCRP member center in Georgia — had its YouTube channel taken down last week over three bogus copyright claims, on videos about a politician's wife's Hermès collection, sanctioned judges, and a tax break for a government-linked businessman. A reporter emailed the complainant posing as a client, and within minutes had a WhatsApp number and a price list: $3,000–$5,000 to nuke a channel, $380 a video to walk a claim back once caught.

The same racket has hit Radio Azattyq (RFE/RL's Kazakh service) across Instagram, YouTube, and Facebook, flagged over stories on a presidential decree and a politically connected company. A researcher at Citizen Lab makes the point that matters here: this kind of takedown is invisible to everyone except the outlet losing its channel, and the burden of proof sits entirely on the victim.

The Wall Street Journal published a piece last night that's worth your time if you do any work involving US government requests to platforms. It opens with an anonymous Reddit account, "Tired_Thumb," who posted 116 comments across 40 subreddits between January and mid-February, three of them critical of ICE. By March, that person was in a constitutional fight after DHS tried to unmask them.

That's not an outlier case. ICE has built a 24/7 operation scanning Facebook, Instagram, and X for anything that might "endanger the agency's mission," paying private contractors to turn public posts into dossiers with a target's name, location, date of birth, employer, and vehicle records. To unmask anonymous accounts, DHS has sent hundreds of subpoenas to social platforms, and reporting says they've recently escalated to grand jury subpoenas, which are harder to challenge than administrative ones.

Worth reading closely if you build dossiers from public posts for a living. It's the same open-source method, pointed at people documenting or criticizing a government agency instead of at bad actors.

OpenAI quietly extended SynthID watermarking to audio this week — anything generated through GPT-Live (ChatGPT Voice, the API) now carries the same invisible watermark that's been on their images since May. The bigger deal is what shipped alongside it: API access to the detector, so you can check a file programmatically instead of uploading one clip at a time to their web tool. That's the difference between a party trick and something you can actually build a triage step around.

Worth knowing the limits before you rely on it. This only flags OpenAI-origin audio — anything from ElevenLabs, an open-source voice clone, or literally any other generator comes back with no signal, and "no signal" doesn't mean authentic, it just means no OpenAI watermark was found. So it's a useful "yes, this is confirmed as ours" check, not a general deepfake detector. Also worth clocking the timing: this landed two days before the EU AI Act's Article 50 disclosure rules kick in, which is probably not a coincidence.

If audio verification is anywhere in your workflow, this is worth a look now rather than filing away for later — it's the kind of thing that's more useful once you've actually run a real file through it and know what the output looks like.

Google Earth carries default authority. It gets used to confirm timestamps, geolocate footage, and check infrastructure changes in conflict and disaster reporting. Once fabricated content sits inside the same interface, using the same coordinates and camera controls, “verified satellite imagery” and “AI-generated content” become visually indistinguishable to anyone viewing a screenshot.

The geometry is also real: correct shadow angles, correct building footprints, correct terrain. Shadow analysis and structural continuity checks, two of the standard tells for a fabricated image, rely on geometric implausibility. Here only the surface content is synthetic, which is precisely the kind of forgery those checks are not built to catch.

Google has built in watermarking, visible and invisible, and blocks direct export or download from the platform. Neither protection survives a screenshot. On-platform labelling and export blocks do nothing once an image is captured and reposted, which is the default path for anything that ends up in a feed.

The historical reconstruction use case is the one to watch. The same workflow that renders ancient Pompeii will render a fabricated “before” image for a contested site, a conflict zone, or disaster damage. Nothing in the announcement signals restrictions on that.

This ships globally, in-browser, no access tier, no stated use-case gating. A mass-market fabrication tool layered onto the world’s most recognisable geospatial reference product.

If yt-dlp sits anywhere in your archiving pipeline, go and read the 2026.07.04 release notes before your next run. Two things will bite you. First, --exec now only accepts safe conversions in command templates, so anything using %(...)s needs to become %(...)q. Numeric conversions are unaffected. Second, aria2c support for HLS and DASH downloads has been removed outright — if you were leaning on it, switch to -N for concurrent fragments via the native downloader.

The security side is worth a minute too. CVE-2026-55404 is a downstream command injection through unsanitised --write-link output, and the previous release fixed a cookie leak when curl was the downloader plus arbitrary code execution via aria2c manifest downloads. Anyone pointing yt-dlp at adversarial URLs — which is most of us, most of the time — is exactly the threat model these describe. Python 3.10 support is also on the way out.

Slightly outside the usual two-week window at 24 days, but I'd rather flag it late than not at all, since broken --exec scripts tend to fail quietly rather than loudly.

The Insider went through customs records and found that every bit of beryllium oxide ceramic reaching Russian defence plants comes from a single Kazakh manufacturer, Kaz Ceramics, sold either directly or through an affiliate called Zenit-K. Russia sits on some of the world's largest beryllium reserves and has never developed them, so it imports. Before 2022 that was about $1m a year. In 2025 it was around $8m.

What makes this one useful rather than just interesting is the shape of the method. Andrey Zayakin identifies a material with an unusually narrow use case (it's a heat-dissipating dielectric, so it ends up in radar transistors and missile seeker heads), traces who ships it, then lists the receiving plants one by one against sanctions designations. Most of the buyers are already sanctioned. A couple, including Fazotron-VMZ, which makes an emitter that shields Iskander-M missiles from air defence, are not. That gap is the story.

There's also a nice wrinkle at the end: Kazakh prosecutors did go after Kaz Ceramics, but over shipments to Belarus, not Russia. The court effectively classified the ceramics as dual-use goods. So the legal record already contains the finding that the reporting needed.