All Posts

A paper published today — Stealing Reasoning Traces from Proprietary LLM APIs, arXiv:2608.09867, from a team spanning the ELLIS Institute Tübingen, the Max Planck Institute, MATS Research and Snyk — does something more useful than demonstrate a clever attack. It measures the damage on real data. The researchers took 6,708 publicly published agent trajectories, decoded 315,320 encrypted thinking blocks, and recovered 704 distinct privacy artifacts. Among them: 62 API keys, 33

Alejandro Hernando and Borja Martínez presented "Plug & Pwn" at DEF CON 34. The PoC is now public. The short version is that you can plug a device into a fully patched Windows 11 machine with nobody logged in, and end up with code running as NT AUTHORITY\SYSTEM — without a single click, and without exploiting a vulnerability in Windows. We are covering it because it belongs to a class we have been tracking all day, and because the framing in most of the coverage undersells wh

We have been covering the residential proxy racket for five months, so we want to put three dated events next to each other, because together they describe a supply chain under pressure and an acquisition method moving. March 12, 2026. The FBI issues advisory PSA260312 on criminal and nation-state abuse of residential proxy networks, without naming one. We published 1,360 indicators against it on April 28. July 2, 2026. Google's Threat Intelligence Group, the FBI, the IRS and

Reco published research this week on a campaign they call City-Forum, named for a domain tied to the attacker's IP. A single actor has been systematically reading data out of Salesforce Experience Cloud sites and ServiceNow Service Portals around the world since at least March 2025. Seventeen months. Telecommunications, financial services, enterprise software, security and data-privacy vendors, public-sector portals. All of it from one IP address: 158.220.87.79, a Contabo VPS

Microsoft shipped its August Patch Tuesday on the 11th. Cisco published an advisory the same day. CISA added three CVEs to the Known Exploited Vulnerabilities catalog on the same date. Most of the coverage counted the CVEs and moved on. We ran the batch against our own KEV index instead, and two of the three turned out to be sitting on attack surfaces that are among the most repeatedly exploited in the catalog — surfaces that do not appear on anyone's "most exploited products

Our morning sweep flagged one real gap today: DeadLock ransomware, which stores its command-and-control configuration in Polygon smart contracts. Six outlets ran it in the last forty-eight hours off a Microsoft writeup dated August 10. We had no coverage. So we went to write it, and the research produced a better story than the gap did — one that costs us something. The Thing We Got Wrong On July 16 we published a post about Starland RAT, a credential stealer that keeps a bac

On August 10 the FBI, CISA, NSA and South Korea's National Police Agency published a joint #StopRansomware advisory on Gunra — AA26-222A. Gunra emerged in April 2025, is believed to derive from the leaked Conti source, has grown into a full ransomware-as-a-service operation with an affiliate programme, and has listed 51 victims across healthcare, financial services, critical manufacturing, transportation and logistics, government, and utilities. It is a good advisory. It is a

The Register reported on August 10 that advertisers are placing hidden content aimed at AI crawlers, and that Time Magazine is serving brand material in markdown versions of its pages that ordinary readers never see. We do not repeat claims we can test. So we tested it, eleven different ways, and the result is more specific and more interesting than the headline. The measurement We requested https://time.com/ eleven times on August 11, changing only the User-Agent header. Sam

We have a standing beat here about vulnerabilities with no CVE: remote code execution quietly fixed in a release note, never filed, invisible to every scanner keyed on identifiers. The label failure is the story, because a bug your tooling cannot name is a bug your tooling cannot find. TrueConf is that beat escalated one level, and it is worse in an instructive way. The identifiers exist. They are in a registry your tooling does not speak. Two attacks, one product, opposite d

A family of malicious Visual Studio Code extensions published under the name Solidity Pro has been stealing crypto wallets, seed phrases, cloud credentials and source-control tokens from blockchain developers. Yeeth Security documented the current wave on August 6; the reporting was picked up widely on August 10. Most of the coverage describes it as an infostealer in an IDE extension, which is true and is the least interesting thing about it. Three design decisions in this th

North Carolina State Ports Authority was attacked on August 4. The disruption reached all three of its facilities: the deepwater terminals at Wilmington and Morehead City, and the inland terminal at Charlotte. The response, on the public record, was fast and unglamorous. The IT team activated a Cybersecurity Contingency Plan. Wilmington ran a delayed opening and switched the truck gates to manual processing, deliberately, so that the IT side could concentrate on recovery inst

Novee Security presented four sessions at Black Hat USA and DEF CON last week. Three of them were the same finding in three different products: Anthropic's Claude Code, Google's Gemini CLI, and OpenAI's Codex. Elad Meged and colleagues showed that a stranger with no repository privileges could open a public GitHub issue and take over the automation that was supposed to be maintaining the repository. Two CVEs came out of it. CVE-2026-54316 in Claude Code, CVSS 9.1 at NVD, ever

Somebody calling themselves AMOZIHEV took a run at Colibri Group, and they did not do any of the things a ransomware crew is supposed to do. There is no leak site. There is no onion address. There is no countdown timer on a dark-web portal that four hundred threat-intel analysts screenshot and nobody else ever sees. There is no post on a breach forum. What there is, instead, is a notice addressed to Colibri Group, published on Colibri Group's own production websites, in front

On July 30 we published a piece about CVE-2026-63077, an unauthenticated remote code execution bug in the protocol JetBrains TeamCity build agents use to phone home. CVSS 9.8. We had searched GitHub that night and found no public proof-of-concept; our exploit harvester held nothing either. The advice was to patch immediately, on the grounds that the quiet week before someone publishes an exploit is the cheapest time a defender ever gets. The quiet week lasted six days. On Aug

On August 6, OpenSourceMalware published a campaign of nearly 800 malicious npm packages delivering a cross-platform RAT and infostealer through a downloader they named WEL1DROPPER. Sonatype tracks the wider wave as sonatype-2026-005660 and counts 846 components. The Hacker News and SC Media picked it up the following day. The reporting gives defenders a list of package names. Lists of package names are the least durable thing you can be handed, because the next campaign uses