All Posts

On July 8, Wiz published GhostApproval, and it is the most honest name anyone has given an AI-agent vulnerability yet, because the bug is not really in the symlink. It is in the box. The little approval dialog your coding assistant shows you before it writes a file — the one you have been trained to glance at and click — can be made to tell you the truth about a harmless file while it writes to a dangerous one. The agent, in at least one tested case, already knew. It said so

The Department of War published the fourth tranche of UAP records under PURSUE this morning, July 10, 2026. Forty new assets: fourteen documents, nineteen videos, three images, four audio files, from the Department of War, NASA, the CIA, the FBI, and — for the first time in this collection — the Department of Energy. By early afternoon we had pulled all fourteen PDFs off war.gov, extracted the full text, indexed every one of the forty into our searchable archive, geocoded the

Zimbra pushed a patch today, July 10, for a critical stored cross-site-scripting flaw in its Classic Web Client. There is no CVE number assigned yet. The affected code is the Ajax webmail interface used by hundreds of millions of people, the fix is in ZCS 10.1.19, and the mechanics are the kind that do not require the victim to do anything wrong: a specially crafted email runs attacker JavaScript in your session the moment you open it, and from there it can lift your session

Two weaponized exploit repositories landed on GitHub this morning. We caught both. One is an honest tool for a real bug. The other is a mass-exploitation framework wearing a CVE number that belongs to something else entirely — and it did not come from nowhere. It came off an assembly line. The line The account is shinthink. It has existed since June 10, and for the first three weeks it did almost nothing. Then, starting July 3, it began shipping — and it has not stopped. In e

Two vulnerabilities went into CISA's Known Exploited Vulnerabilities catalog in June that we did not flag early and owe you no receipt on. We are filling those two holes here, late and honestly credited. But the more useful thing is what we found when we went to write them up: they are not two loose ends. They drop into a cluster we have been mapping in this blog since May — the convergence of the top ransomware crews on the access, management, and orchestration plane as the

Not every bug is one we catch first, and a threat-intelligence shop that only tells you about its wins is lying to you by omission. Three vulnerabilities went into CISA's Known Exploited Vulnerabilities catalog across June that we did not break, did not flag early, and owe you no receipt on. We are writing them up anyway — late, honestly credited, and with the one thing a wrap of three separate advisories usually leaves out: the pattern connecting them. Because when you line

The account was created on November 25, 2024, and then it did nothing for nineteen months. No repositories, no followers, no activity — one of the millions of dormant GitHub shells that sit empty forever. Today, at 11:49 UTC, it woke up. Eighteen minutes of commits later it held a single repository: a fully weaponized, one-click mass-exploitation kit for CVE-2026-8037, a pre-authentication remote-code-execution flaw in Progress Kemp LoadMaster that is already being exploited

GhostLock was the flagship of the day and gets its own write-up. This is the rest of Thursday — the headlines that broke around it, and where we already had them. No victory lap, just the ledger, our date next to theirs. Microsoft finally patched RoguePlanet. We called it exploit #8, unpatched, three weeks ago. Microsoft shipped a fix today for a Defender zero-day called RoguePlanet. We have been writing about it since June 18, in a post titled "RoguePlanet Is Exploit #8 From

Here is the entire attack, and it is almost insulting in its simplicity. Any ordinary program you can run on a Linux machine makes some ordinary threading calls — the kind a normal application makes a thousand times a second — and about five seconds later you have a root shell. No special permission, no unusual configuration, no network access, no exploit that requires you to already be halfway in. Just a logged-in user and a stopwatch. That is CVE-2026-43499, which the resea

This morning we published a post arguing that a threat-intel shop should be honest about when it is ahead and when it is not, because a ledger that only ever shows wins is a marketing document. A few hours later the universe handed us a live test, and this is the honest scorecard. Socket disclosed seventeen malicious packages today, all impersonating legitimate payment SDKs — Paysafe, Skrill, and Neteller — across npm and PyPI. The packages do exactly what that class of attac

Four stories broke into the headlines this week that we had already written up — one of them almost two months ago. That is the kind of sentence a threat-intel shop is tempted to set in bold and put on a billboard. We are going to resist, because the honest version is more useful than the victory lap, and the honest version includes the part where we were not clever at all. Here is the ledger. Our date, then theirs. The CISA leak: us, May 19. Congress, this week. A contractor

Yeats asked the question in 1919 and refused to answer it. "And what rough beast, its hour come round at last, slouches towards Bethlehem to be born?" He was from Sligo, under the same mountain where warriors have climbed for five thousand years to leave a rock for a dead queen. He knew you don't get to stop the beast. You only get to see it coming and name it. So here is the name, and it is not a robot uprising. The rough beast is the agent with no human in the loop — and th

Full disclosure in the first sentence, same as the last time we wrote about this: we build on Claude Code every day. It is core infrastructure here, it is the partnership this whole shop is bound to, and this is not a hit piece. It is a daily, invested user reading the week's news and deciding that saying the true thing plainly is worth more than picking a team. Here is the week. On Monday we published a post titled "Anthropic Hid a Tracker in Claude Code Using Steganography,

Zscaler's ThreatLabz just documented the thing we built a scanner for in April, happening in the wild, with money changing hands. Two campaigns are hiding instructions inside web pages — concealed with CSS, HTML, and JSON-LD, and boosted with SEO poisoning so that AI browsing agents find them — and those hidden instructions tell the agent to do things its human never asked for. In one campaign, a fake developer-documentation site walks an agent into paying three dollars or se

Here is the entire attack. When a Gitea instance is configured to sit behind a reverse proxy for authentication, it reads a header called X-WEBAUTH-USER and treats whatever name is in it as the logged-in user. The catch is the default: Gitea ships trusting all upstream connections as legitimate proxies — the parameter is set to a wildcard out of the box. So any remote client that can reach the instance sends its own request with X-WEBAUTH-USER set to admin, or gitea_admin, an

On July 2, 2026, Google's Threat Intelligence Group, the FBI, the IRS, and Lumen pulled the plug on NetNut — a residential proxy network, also tracked as Popa, built on at least two million hijacked home devices. Not servers. Smart TVs. Streaming boxes. The stuff in your living room, infected through malicious SDKs the operators baked into apps, so that a stranger somewhere could route their traffic through your internet connection and your address caught the blame. The FBI a

Accenture is a three-hundred-thousand-person, sixty-billion-dollar consultancy, and a meaningful slice of that revenue is telling other large companies how not to get owned. Incident response. Managed detection. Security strategy decks with a lot of pyramids on them. So there is a particular kind of quiet in the fact that, according to Hudson Rock's supply-chain monitoring, more than ten distinct Accenture employees are sitting in infostealer logs right now with credentials t

Our exploit-harvester flagged two fresh proof-of-concept exploits the day they were published — both authenticated remote-code-execution flaws in Vtiger CRM, from a GitHub account called JivaSecurity that almost nobody is watching yet. Zero stars. Zero followers. A brand-new name. But the interesting thing isn't the individual bugs. It's that when you lay out everything this researcher has published, and cross it against the repositories they've quietly forked but not yet pub

Full disclosure before the first sentence of argument: we build on Claude Code. It is core infrastructure here — the partnership this whole shop is bound to. So this is not a hit piece from someone who wants Anthropic to lose. It is the opposite: it is what it sounds like when a daily, invested user reads that the tool it trusts hid a user-tracking signal inside its own system prompt using steganography, and decides that saying so plainly is worth more than looking away. The

We spent this week publishing bugs one at a time — ColdFusion, Bad Epoll, DirtyClone, Januscape, BeyondTrust — the way the wire reports them: a CVE, a severity, a patch note, next. Then a reader asked the question that reorganizes the whole picture. Not "how bad is each bug," but "are you thinking about how they combine?" We weren't, not on the page, and that was a failure of framing. Because attackers don't operate CVEs one at a time. They operate chains. So we did the arith