The Briefing

What happened, why it matters, and what to do about it. Analysis, not rewritten press releases.

Showing 241 to 264 of 658 briefings

Tuesday 6 October

AI Security6 Oct 2026

Anthropic's own test: Red Team access blocked nothing, as it expected, and Defense let 4 of 50 trials succeed

Anthropic's 6 October CVP page: Red Team 34 of 50 against 67.6% unfiltered, Defense 4 of 50 through. What the run does not show, the 134,500 against 135,610 vulnerability counts, UK GDPR and Computer Misuse Act points, and a checklist.

Market and Supply Chain6 Oct 2026

Outlook's .msix block adds two extensions to a 133-entry list, in two clients, with no abuse figure

What Microsoft's notice states about the .msix and .msixbundle block, what it leaves out, the routes it does not touch, and a UK checklist for Exchange Online administrators before the early-November rollout.

Threat Intel6 Oct 2026

Passkeys urged after a 417% rise in reported losses that Report Fraud's own assessment links to its new service

Reports rose 34 per cent but victims reporting a loss rose 10.3 times, and no document says how the accounts were taken. A passkey closes one door; recovery and fallback sign-in are others.

Breaches and Incidents6 Oct 2026

ASOS confirms the 10am notification was unauthorised, cites third-party platforms and does not name Snowflake

The channel is confirmed and the Snowflake claim is not addressed. What the RNS says and leaves out, what UK GDPR and UK MAR say, and what to do about every platform that can message your customers.

Threat Intel6 Oct 2026

LibreOffice's fix for a spreadsheet code-execution flaw came 73 days before its advisory; OpenOffice has none

The LibreOffice fix shipped on 24 July and the advisory on 5 October. OpenOffice still has no released fix and a public proof of concept has existed since 5 October. What each project documents, and what to do first.

Market and Supply Chain6 Oct 2026

Forescout's 6% for medical devices is an OpenSSH version test, not proof they cannot be upgraded

What Forescout's report states about 6% of medical devices and 31% TLS 1.3 on exposed systems, what it leaves out, what the NCSC, NHS England and the MHRA say, and a UK checklist in the order worth doing.

AI Security6 Oct 2026

Adversa's Copilot CLI attack ran in 50% of runs on one of three models, and the post gives no run count

One model, 50% of runs, no run count: what Adversa's Copilot CLI write-up shows, what it leaves out, what GitHub's own pages say about autopilot, and what UK teams should do first.

Market and Supply Chain6 Oct 2026

One Debian kernel update lists 1,313 CVEs, 97% of the previous 17 combined. None is on CISA's KEV list

The count reflects how the kernel's CNA assigns identifiers, not how much danger one update removes. For UK estates the Cyber Essentials 14 day clock runs to 13 October 2026.

Compliance and Regulation6 Oct 2026

OpenAI's own test: its text watermark falls from about 92% to 17% detection when 25% of words are replaced

In OpenAI's own test, its textGrain watermark was found in about 92% of unedited 400-token passages and 17% after 25% of words were replaced. What that means for the EU rule behind it and for UK universities, employers and publishers.

Threat Intel6 Oct 2026

Seven of Atlassian's 19 fixed builds for CVE-2026-21589 are on lines that leave support by 30 December

Eight Atlassian Data Center products, 19 fixed builds, a 9.3 that computes to 8.7 without its downstream assumptions, and a CVE record whose Crowd and Bamboo builds disagree with the bulletin. What the sources state, and what they do not.

AI Security6 Oct 2026

Wikimedia lists 54 edits it attributes to OpenAI agents. 49 are on sandbox pages, and its method is not stated

Wikimedia's 5 October post links 54 edits it attributes to OpenAI agents: 49 sandbox pages and five citation-tool settings. It reports no compromise and does not state how it made the attribution.

Breaches and Incidents6 Oct 2026

Bromcom found a legacy sign-on breach on 6 September; its first dated public post came 18 days later

A superseded sign-on service stayed in production because an internal system still called it. Bromcom found the breach on 6 September, posted on 24 September, and by 30 September called the central question ongoing.

Page 22

Monday 5 October

Market and Supply Chain5 Oct 2026

Cloudflare's post-quantum CA has issued nothing yet, and it targets authentication, not recorded traffic

What Cloudflare, Chrome, Apple, Microsoft, Mozilla, the IETF and the NCSC have said about post-quantum certificates, what is still a draft, what the GlobalSign deal covers, and a UK checklist in the order worth doing.

Breaches and Incidents5 Oct 2026

Microsoft's X account: two unauthorised posts, 47 minutes apart by our arithmetic, and no word on the cause

Microsoft's 13-million-follower X account carried a disclaimer that told readers to trust only official channels, and Microsoft says it did not write it. The minutes, what is not stated, five routes in, and a UK checklist.

Threat Intel5 Oct 2026

$40.73M over 1,500+ ATM attacks is at most about $27,000 each, and the charged route starts at the cabinet door

Divide Treasury's headline loss by its own attack count and the average is at most about $27,000. The DOJ, FBI and ATM makers all put the real control at the cabinet door, the keys, the alarm response and the disk, not at the arrest.

Threat Intel5 Oct 2026

A Samsung signage flaw at least 476 days old led to AnyDesk on try three and a miner built on the host

At least 476 days between the CVE record and the incident, three AnyDesk download attempts, eight days between two reports on one endpoint. What Huntress states, what it does not, and the order of checks for signage servers.

AI Security5 Oct 2026

OpenAI to test ads during ChatGPT image generation in the US; its free no-ads option has no image tools

OpenAI will test ads during ChatGPT image generation, in the US, later in October. A read of its ads pages and policies, and of ICO, ASA and CMA guidance, sets out what is stated, what is not, and what a UK organisation should check.

Threat Intel5 Oct 2026

Cling botnet poses as STUN traffic, but gets in through flaws catalogued up to 4,175 days ago

A Linux botnet dresses its command traffic as STUN, the NAT-traversal protocol behind video calls, but it gets in through flaws first recorded between 2015 and 2026. The disguise is the news; the doors are the part a UK security lead can still control.

Threat Intel5 Oct 2026

Dell posted the DSU fix 65 days before it called the flaw critical, and the download says Optional

The lead flaw is scored Network with no privileges, so it is not a local escalation by Dell's own scoring, but all five need user interaction Dell never explains. One fix, no workaround, nothing on exploitation.

Threat Intel5 Oct 2026

Android 17's accessibility lock works only if the user turns Advanced Protection on; no uptake figure found

Google's accessibility restriction needs three things at once: Android 17, Advanced Protection switched on, and an app that is not a verified accessibility tool. Only the first has a published figure, and it is a third party's estimate.

Breaches and Incidents5 Oct 2026

About 9,000 spoofed Nikkei emails reached news sources among others; who was exposed is not yet counted

What Nikkei's notices say in Japanese and English, what they leave out, how Japanese and UK law treat the security of data held for journalism, and an ordered checklist for any organisation that holds the names of sources or whistleblowers.

Threat Intel5 Oct 2026

Cache key injection has a 2020 name, four 2026 lab demonstrations, no CVE and no named victim

A research technique, not an exploited flaw: what the write-up states, how Kettle's 2020 paper named the same bug class, what five vendors' documents say about how keys are joined, and a UK checklist for cached personalised pages.

Breaches and Incidents5 Oct 2026

Frontline's breach letters were first reported 48 days after it found the flaw, and name no software

A supplier to US school districts found a flaw on 14 August and was first reported notifying districts on 1 October, 48 days later. Staff Social Security numbers were taken; the product, the counts and the actor are not stated.

Breaches and Incidents5 Oct 2026

Denmark says one company's lawful access reached 8.8 million of the 11 million entries in its CPR register

About 80 per cent of the register, more than the population, reached through one company's lawful access. What the ministry and Datatilsynet state, what CPR's own terms say about controls, and the checks a UK team can run.