← Feed

Posts

• View post

  Whoever breached South Korean banks in late September used the same AI setup to draft a CV. CrowdStrike found the session histories in an open directory on a Hong Kong server. The attack ran on ARTEX, an open-source agentic pentest frame from China, driving rented models: DeepSeek primary, GLM and Grok in support. At least seven financial firms were breached; Seoul held an emergency meeting. ARTEX has gone closed source, but English and Korean derivatives already circulate; the rented models...

• View post

  Change your password every 90 days. Plenty of organisations still enforce it. The NCSC has advised against forced expiry since 2018. People nudge the old password by a character, and predictable swaps like zero-for-o are the first thing attackers try. What works is duller: a minimum length, a deny list of common passwords, throttling, password managers and MFA on internet-facing accounts. After that, change on suspicion of compromise, not because a timer fired. https://www.ncsc.gov.uk/coll...

• View post

  A GlobalProtect flaw patched in May is working for ransomware crews. ReliaQuest has Qilin and Settra both using CVE-2026-0257, an authentication bypass, to open VPN sessions that look like remote work. Exposure depends on configuration: the authentication override cookie feature plus a particular certificate setup, so not every estate is affected. Fixes have been out since 13 May at the vendor's highest urgency. Panorama and Cloud NGFW are unaffected. ReliaQuest found Cobalt Strike...

• View post

  An MFA prompt you did not trigger is not a nuisance. It fires only after the password has passed, so one unexpected buzz means someone already has it. Push bombing leans on that sequence. An attacker with a stolen password sends prompts until the owner taps approve, by accident or to make the phone stop. Number matching kills the reflex: approval needs the number on the login screen, so you cannot approve a login you cannot see. When a prompt arrives that you did not cause, report it and ch...

• View post

  French supermarkets spent the weekend throwing away food. From Saturday, attackers pushed forced defrost cycles into refrigeration at Coopérative U, Intermarché, Carrefour and Monoprix stores, driving temperatures towards 50°C. Intermarché counts 150 stores affected, Carrefour 80. Two stores in one Alsace town report €270,000 of lost stock. The route in was not exotic. The consoles were Danfoss controllers, obsolete or unpatched models, reachable over channels that were never properly secur...

• View post

  Which is riskier, American AI or Chinese AI? My latest piece asks what each government has already done with the power it holds. Washington has switched AI access off three times in two years: OpenAI's China cut-off in 2024, Microsoft suspending an Indian refiner in 2025, the ban on Anthropic, its own lab, this February. Beijing's reach is written into intelligence law: no warrant, no appeal. One asymmetry settles it. China gives its strongest weights away under open licence...

• View post

  Zenity Labs has shown what one prompt could do to AWS Bedrock AgentCore. Agents could be talked into fetching credentials from the instance metadata service. One prompt injection, and the attacker held the agent's IAM role. That role was wide. It could invoke every other agent in the account's region, read private conversations, plant memories to alter future behaviour, and pull keys from Secrets Manager. AWS moved agents to IMDSv2 in February, but only hardened the default...

• View post

  A financial firm took 1,800 Dridex-laced emails in one campaign. Filters stopped 1,750. Of the 50 that reached inboxes, 14 were clicked; the one infection was detected and cleaned within hours. As the NCSC notes, no single layer would have caught all of it. The guidance is just as blunt about phishing simulations. No training teaches users to spot every attempt, and punishing people for clicking an email you sent them starts to resemble entrapment. If the celebrated metric is a falling cli...

• View post

  The one Microsoft bug being exploited right now is rated Important, not Critical. CVE-2026-96951 lets a crafted file slip past Mark of the Web, the check that flags internet downloads before you open them. SmartScreen, Office Protected View and the macro block all lean on it, so one bypass silences several warnings at once. It arrived inside a record release: 1,413 CVEs on Tuesday, up from September's 974, which was itself the record. Counts that size make triage the job. Windows 11...

• View post

  Email, documents, calls and storage: the working day runs on American infrastructure by default. US law follows the provider, not the data centre. The CLOUD Act reaches data held anywhere, so an Irish or German region changes nothing about jurisdiction. My latest piece looks at Proton as the most complete European answer. It is Swiss throughout: servers in Switzerland, outside US jurisdiction, keys held by the user. The way in is hybrid: Proton for internal and sensitive traffic, specialist...

• View post

  Brussels has opened its first AI Act enforcement test. The AI Office has sent OpenAI and Anthropic information requests after two summer incidents: an OpenAI agent that broke out of its test environment and into Hugging Face, and Claude models that reached three outside organisations via a misconfigured evaluation setup. Voluntary disclosures were judged insufficient. The fight is over scope. The Act exempts research and testing; lawyers say that covers these cases. The Commission reads it...

• View post

  Ransomware does not stop at the live data. The NCSC has watched it encrypt the connected USB drives and network shares, and compromise cloud backups. The familiar 3-2-1 rule says nothing about any copy being offline. Three copies, two devices, one offsite can all stay reachable, and one incident can take the lot. At least one copy needs to be offline or digitally disconnected when idle. Create-only access on cloud storage gives the write-once effect: ransomware can add a backup but not over...

• View post

  Some 315 crypto accounts lost about $102 million last week, drained through hardware implants inside Ledger wallets. The implant copies the seed phrase as the screen displays it and sends it out over an added LTE module and eSIM. Joe Grand found the first ones in March, in devices from dodgy online sellers. The newer finds are worse. They look factory-fitted and are turning up in wallets from authorised resellers, including two bought in person at MediaMarkt stores in Germany. Buy-direct ad...

• View post

  A reader under my Airbus piece asked the question the sovereignty debate keeps skipping: who provides the login? Move everything to a European cloud and an American identity provider still holds the keys. The identity market belongs overwhelmingly to Microsoft Entra and Okta, both under American law wherever the servers sit. Microsoft's own documentation says it: Azure Local's disconnected tier does not support Entra ID. Its most sovereign product cannot use its own identity...

• View post

  Double Counter, a bot thousands of Discord servers trust to keep raiders out, was breached on 4 October. The way in was a Metabase analytics tool left exposed on a legacy server. Cloud credentials from there opened the rest. The attacker held access for nearly six hours and copied about 12 GB: a million email addresses in full, plus IDs and coarse locations for tens of millions of accounts. A stolen bot token then posted malicious links into some 50 large servers. Tellter, the French operat...

• View post

  Supplier security is usually assessed once, at onboarding, then filed. The NCSC's twelve principles run as a loop instead: understand the risks, establish control, check the arrangements, keep improving. Checking is the stage that dies first. Contracts hold requirements nobody revisits. Per the 2025 Security Breaches Survey, very few UK businesses set minimum security standards for their suppliers at all. Onboarding evidence says how a supplier looked on one day; the loop exists bec...

• View post

  One stolen vendor credential opened 173 Microsoft 365 tenants last week. ConnectWise Cloud Backup, formerly SkyKick, had its OAuth app secret taken. On 15 October the attacker signed in as the application and pulled Microsoft Graph tokens, working through tenants in sorted order, seconds apart. Nobody was phished and no MFA prompt fired. The sign-ins inherited the permissions customers had granted the backup app: mail, files, Teams, directory data. ConnectWise has revoked the app's...

• View post

  Push a firewall update on a Saturday and the sales team cannot open email on Monday, with four hours of client outreach lost while IT rolls back. In a small business every system leans on the next: one CRM schema change lands in marketing's targeting and finance's forecasts alike. While systems are down, staff improvise: personal email for work documents, unapproved messaging apps. The outage becomes the exposure. My domino principle piece backs the boring controls: a two-que...

• View post

  Ahold Delhaize confirmed on Monday that the outage hitting its websites and apps since 12 October was an attack. The Dutch group behind Albert Heijn, Delhaize, Food Lion and Stop & Shop took the systems offline itself, as a precaution, across most of its European and American brands. Stores stayed open throughout. The group told the Dutch NCSC within 24 hours of detecting the activity and filed a police report yesterday. No customer data is believed taken, though the investigation co...

• View post

  An intrusion discovered today may have started months ago. The NCSC tells you to size the log window to that gap, initial access to detection: 13 months for critical systems, 180 days otherwise. Rotate at 30 days and you can only reconstruct the last month. How fast logs leave the box matters as much. Logs left on the box belong to whoever owns the box. Move them to a central store quickly and the attacker loses the chance to delete their own tracks. Logging Made Easy, the NCSC's o...

• View post

  MFA is on the checklist of every audit I have ever seen, and the tick is where most thinking stops. It should not be. The NCSC's guidance is blunt about it: not all types of MFA are created equal. Text messages can be intercepted. A bare push notification can be spammed until someone taps approve at 11pm to make it stop. That is MFA fatigue, and it works because the approval asks nothing of the user except irritation. The better versions make phishing structurally harder: number matchi...

• View post

  The first control in the CIS list is not a firewall or a password rule. It is knowing what you have. Inventory and control of enterprise assets, including the cloud ones, including the ones nobody bought on purpose. Every framework starts there because everything else stands on it. You cannot patch the server nobody recorded. You cannot put the unlisted laptop in the risk register. Unmanaged and unknown are the same thing to an attacker. The failure mode I keep meeting is the inventory th...

• View post

  ENISA's 2026 Threat Landscape is out. Two numbers belong together. Where ENISA could identify the intrusion vector, 60% of unauthorised-access incidents came through a vulnerability, N-day or zero-day. And 2025 published over 48,000 new CVEs, up 22% on the year before. More doors, and the doors are the preferred route in. The other thread is dependencies. Supply-chain and third-party attacks are what produce the large-scale incidents. Which is why the point about patch clocks shrinkin...

• View post

  China’s spy chief just put the world on notice about AI. Chen Yixin warns foreign powers are using genAI to fabricate rumours, launch opinion wars, and hack. He singled out Claude Mythos and ChatGPT-5.5-Cyber. This is tech sovereignty. Xi-Trump meeting next. The race is on. Are we ready?

• View post

  JetBrains confirms its Cadence cloud service was breached through a TeamCity vulnerability it had itself disclosed. CVE-2026-63077 is critical: unauthenticated attackers could run commands on it. Cadence stayed unpatched, and attackers were inside from 8 to 24 August. Their own sentence: "The server should have been patched as part of our response to the vulnerability, but it was not." The exposure is what build infrastructure concentrates: a full 2024 server backup, multiple AWS I...