The Global Threat Summary, week ending 24 October 2026

In one line: the software under most of the world's websites needed an urgent fix, the connections and certificates organisations trust were the way in, and three separate investigations showed attackers using AI agents to do the work.

The first is the web itself. On 21 October PHP, the language behind a large share of the world's websites, fixed two critical flaws that let someone with no login run code on the server. Attempts to exploit them began within hours. Governments in the UK, Canada, France, Finland and Australia took websites offline as a precaution, and ASD's own cyber.gov.au was caught in an outage. One security vendor reports a break-in at a European government body.

The second is trust. Microsoft says a China state aligned actor was behind last week's compromise of a backup app trusted inside Microsoft 365 tenants. Eight certificates from an Indian certificate authority turned out to have private keys anyone could work out, so browsers and Cloudflare raced to stop them being used to impersonate websites. Canva confirmed that a connected feedback tool exposed some enterprise customer records.

The third is cost. Microsoft, Palo Alto Networks and an independent researcher each describe attackers running AI agents. In one account, a single operator tried more than 1,200 websites for about US$5,000 in AI fees. Separately, US agencies say SafePay ransomware has hit more than 600 critical infrastructure providers, with Australia one of its focus regions.

Three questions for the executive team.

1. Who runs our website and web applications, which version of PHP is on each today, and who confirmed it?
2. Which apps, connectors and vendors hold standing access to our email, customer records and cloud, and how quickly could we switch one off?
3. Attacks are getting cheaper to run. When did we last look at our own business from the outside, the way an attacker would, and what did we fix?

Cross sectoral, the items that reach every industry

The web's foundations

PHP2Shell, CVE-2026-109618 and CVE-2026-109619, critical flaws in PHP itself. Cross sectoral. [Confirmed]

• PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35, released 21 October, fix two flaws rated CVSS v4.0 9.2. PHP puts the risk on "all PHP installations serving web requests, including PHP-FPM, Apache with mod_php, php-cgi, FrankenPHP and LiteSpeed", highest for mod_php, and says "There are no complete mitigations, updating is the only safe choice." 8.1 and older "no longer receive updates from the PHP project"; some vendors backport (php.net advisory; PHP 8.5.12).
• Australian exposure, by our own count. GreenCyber's own count from Shodan's index, between 21 and 23 October, of Australian web services that announce their PHP version: 6,923 did. 317 of them, about one in twenty, were on a release that fixes PHP2Shell. 1,636 were on a supported version not yet updated. 4,925 were on a version PHP no longer patches at all. The count sees only servers that announce their version, so the true number exposed could be higher, and distribution backports and the lag in Shodan's index mean some that look unfixed may already be fixed.
• Exploitation. Amazon saw an attempt in its honeypots on 17 October, four days before disclosure, and says "PHP2Shell is now seeing broad exploitation attempts in the wild" and "it's unclear if any of these attempts would have succeeded in real-world systems" (Amazon). watchTowr "has observed exploitation of both" in its honeypots (watchTowr). VulnCheck has published a working exploit, per The Register, read through search.
• One reported break-in, single vendor, lighter. Palo Alto Networks' Unit 42, updated 23 October: "In one case, the attackers targeted a government entity in Europe and within two hours of the initial exploitation they had successfully accessed password hashes and multi-factor authentication secrets." The entity is unnamed and no actor is named (Unit 42). Others still report none: Zafran, "We have no confirmed victims yet" (Zafran); BISI, "millions of exploitation attempts have not produced a publicly confirmed compromise" (BISI).
• Precautionary outages. UK: a minister told the House of Lords 32 government services were unavailable and four partially so, "largely from precautionary measures", and the NCSC sent 50,000 warnings (The Register). Canada's AI Minister said some federal sites were taken down as a precaution (BNN Bloomberg). Australia: Tony Burke's office said "Some websites are temporarily unavailable as work is undertaken to mitigate risks arising from this vulnerability", and ASD said it was "aware of an outage that may be affecting government websites, including cyber.gov.au" (AAP via the Newcastle Herald).
• Divergence on the early release. PHP says Chainguard's early fixed images exposed the flaws before the embargo; Chainguard's chief executive says his company "didn't leak" (The Register). Not in CISA's catalogue at version 2026.10.23. A third, unpatched issue, CVE-2026-115560, needs "an unusual combination of configuration and application code", and PHP suggests disabling session upload progress until its next release (php.net).

Attribution: PHP, Amazon, watchTowr and Unit 42, each read directly; GreenCyber's own Shodan count; the UK and Australian statements as reported by The Register and AAP. ASD's alert, Critical vulnerabilities in PHP, timed out; Cyber Daily reported on 21 October that the ACSC "was unaware of any exploitation" at the time (Cyber Daily).

Eight eMudhra certificates with keys anyone could calculate. Cross sectoral. [Confirmed]

• Cloudflare: "private keys associated with eight eMudhra intermediate certificates were compromised." It learned on 19 October and shipped fixes for WARP, the 1.1.1.1 app and Cloudflare Tunnel on 20 October; "Our security review has not identified active exploitation of this issue on Cloudflare services" (Cloudflare).
• Nothing was stolen. The keys, generated in an April 2025 key ceremony, were "integers below 256 rather than large random numbers", so anyone could forge a certificate for any domain. Firefox and Chrome blocked them on 17 October and Apple on 19 October; Mozilla is removing trust in the emSign ECC Root CA G3 (iTnews; The Next Web).
• Beyond the browser. Cloudflare says customers running their own origin servers, mutual TLS or containers may need to act, and warns against fixing errors by "disabling verification". iTnews found curl on macOS still accepted the chain on 21 October. eMudhra estimates about 30,000 certificates need migrating, per iTnews.
• Network vendors are now publishing advisories: Arista's of 23 October lists the issue as CVE-2026-111943 (Arista advisories).

Attribution: Cloudflare, read directly; browser actions and eMudhra's figure as reported by iTnews and The Next Web.

Fortinet's pre announced batch of critical and high severity fixes. Cross sectoral. [Confirmed]

• Fortinet released a security hardening batch for FortiOS and related products on 21 October, found through its own AI assisted testing: "These releases include critical and high-severity fixes." "Public advisories will be made available at the next Patch Tuesday on November 10" (Fortinet).
• Fortinet names no CVE numbers yet and says nothing of exploitation. Its advice is to patch internet facing devices first.
• Its PSIRT listing showed nothing after 13 October as of 24 October (FortiGuard PSIRT).

Attribution: Fortinet, read directly.

Supply chain and trusted applications

ConnectWise Cloud Backup: Microsoft names a China state aligned actor. Cross sectoral. Since last week. [Confirmed]

• What moved. Last week's Summary said no actor was named. On 21 October Microsoft said: "Microsoft assesses that Storm-0940 was responsible for the recently disclosed ConnectWise supply chain compromise", calling it a "China-state aligned actor" with no known industry aliases (Microsoft). ConnectWise: "it appears that a China state-aligned group was responsible for the attack on ConnectWise Cloud Backup" (ConnectWise advisory; Reuters via KFGO).
• Counts, read carefully. Microsoft says the actor used ARTEX, an open source penetration testing framework that can drive AI models, "to attempt to sign-in to more than 2,250 tenants". That is tenants targeted. Microsoft adds "the majority of the targeting wasn't successful" and that the actor "achieved privileged cloud persistence at many organizations in a brief time", with no count of tenants reached.
• Correction to last week's count. Huntress reported sign in activity from the attacker's address in almost 1,000 Entra tenants, more than 300 of them successful (ITPro). Microsoft: a successful sign in "does not necessarily indicate that customer data was accessed or exfiltrated."
• Restoration. ConnectWise, 23 October, 14:00 UTC: phased restoration, backup restore access first, "We anticipate completing this work by the end of this weekend." Partners keep affected connections disabled until told otherwise. How the credential was first exposed is still under investigation.
• Microsoft lists targeted sectors from aerospace and defence, energy and financial services to government, legal, manufacturing and technology.

Attribution: Microsoft and ConnectWise, read directly; Reuters; Huntress as reported by ITPro. The attribution is Microsoft's assessment, adopted by ConnectWise.

Canva confirms a third party incident through a connected feedback tool. Cross sectoral. [Confirmed]

• Canva, 19 October: "The unauthorised access to Canny allowed access to some enterprise customer records" "through its connection to Canva's Salesforce account." Business contact details and routine documents. "Canva's own platform and systems were not compromised, and Canva accounts, passwords, designs and content remain secure" (Cyber Daily).
• Canva removed Canny's access and informed affected customers where required.
• Claimed. Seven Deadly Sins listed Canva on 16 October and claims more than 2 million Salesforce records, sold to a "private buyer". Canva has not addressed the figure.
• Divergence on the claim. Another outlet reports a claim of more than 200 million database rows (Escudo Digital). Neither figure is verified.

Attribution: Canva's statement to Cyber Daily; the group's figures are claims.

LiteLLM 1.107.0 and the OVHcloud command line tool, both poisoned at release. Cross sectoral. [Reported]

• LiteLLM, a widely used gateway to AI model providers. Version 1.107.0, published to PyPI on 21 October, carried a credential stealer and a remote loader that runs on any Python launch in the environment. PyPI quarantined it within about two hours; 1.106.2 is the last safe version (Microsoft; Wiz; Socket).
• Teams that installed it, or pulled the latest Docker tags that day, should treat the credentials in those environments as exposed.
• OVHcloud. An attacker used a developer's GitHub account to swap releases 0.11.0 to 0.15.0 of OVHcloud's command line tool for a secret stealer on 19 October, for under two hours, after GitHub's release immutability setting was found disabled. OVHcloud says its build pipelines were not affected and advises rotating secrets (BleepingComputer; SecurityWeek).
• Attribution, one source. OpenSourceMalware ties OVHcloud's case to GlassWorm, and says the account was compromised in March (Dark Reading). OVHcloud has not said so.

Attribution: Microsoft, Wiz and Socket on LiteLLM; OVHcloud as reported by BleepingComputer and SecurityWeek. Read through search, except Microsoft's listing.

Exploited at the edge

Arista VeloCloud Orchestrator, CVE-2026-114625, exploited. Cross sectoral. [Confirmed]

• Arista, 22 October: "This issue was detected by Arista and is known to be actively exploited." On premises VeloCloud Orchestrator, the console that runs an SD-WAN. CVSS v3.1 9.9, v4.0 9.4; needs low privileges (Arista advisory 0199).
• Fixed in 5.2.3.18, 6.1.3.10 and 6.4.2.12. Until then, restrict the web interface to trusted management networks and block the two addresses Arista lists.
• Divergence on the flaw type. Arista describes remote access to privileged functionality; CISA's catalogue names it SQL injection. Added 23 October, due 26 October.
• Arista patched a different exploited VeloCloud flaw, CVE-2026-16812, in July, as dated context.

Attribution: Arista, read directly; CISA KEV.

PAN-OS GlobalProtect, CVE-2026-0257: Qilin and INC exploiting a May flaw. Cross sectoral. [Reported]

• Unit 42 reports ransomware groups Qilin and INC exploiting the GlobalProtect authentication bypass, patched on 13 May and catalogued on 29 May as dated context (The Register, 20 October).
• Unauthenticated VPN access, only where a portal or gateway uses authentication override cookies with a vulnerable certificate configuration. Investigations reach back to June.
• Aliases. Unit 42 tracks Qilin as Spikey Scorpius and INC as Tarnished Scorpius, per a German outlet (drweb.de).
• Not the PAN-OS flaw in last week's Summary, CVE-2026-0322.

Attribution: Unit 42 as reported by The Register and German press, read through search; Unit 42's report was not read. Lighter.

Ransomware and state activity

SafePay ransomware: more than 600 critical infrastructure victims, Australia a focus. Cross sectoral. [Confirmed]

• A joint #StopRansomware advisory of 20 October from CISA, the FBI, the Secret Service and international partners says SafePay has struck "more than 600 critical infrastructure providers" since 2024, with "particular focus on North and South America, Europe and Australia" (Cybersecurity Dive; BleepingComputer; CISA bulletin).
• Ways in: compromised credentials, brute force, VPN access and known flaws in internet facing devices, including the Cisco flaws CVE-2020-3259 and CVE-2023-20269. Actors delete shadow copies and reset backup storage.
• Halcyon's Cynthia Kaiser: activity "appeared to have peaked at the end of 2025", and the group has "consistently targeted around 15-30 organizations each month throughout 2026."
• The agencies advise offline, encrypted, immutable backups and multifactor authentication, and note payment does not guarantee recovery. Whether ASD co-sealed the advisory was not established.

Attribution: CISA and the FBI as reported by Cybersecurity Dive and BleepingComputer. The advisory text itself was not read.

AI

Three accounts of attackers handing the work to AI agents. Cross sectoral. [Reported]

• A single operator, more than 100 websites. NorForge's Snorre Fagerland found exposed control panels for Cairn, an open source AI agent platform. One Chinese speaking operator's agents tried more than 1,200 sites from 25 September, across the US, Canada, Australia and Japan, for about US$5,000 in model fees, roughly US$50 per success (iTnews, from Reuters; The Next Web).
• Reuters identified Deere, SharkNinja and Cummins among more than 100 breached, from the attacker's own records. Deere: "no indication that sensitive information was accessed"; SharkNinja says a standalone support site was hit and contained; Cummins is investigating a vendor's website. The models were reportedly DeepSeek and Anthropic's Claude Code. Cairn's author says it "was not designed for hacking attacks." No Australian victim is named.
• A state actor. Microsoft says Storm-0940 used ARTEX to run the ConnectWise campaign, in the trusted applications section above.
• After the break-in. Unit 42 "observed attackers using agentic AI attack platforms for post-exploitation activity" after PHP2Shell (Unit 42).

Attribution: NorForge as reported by Reuters through iTnews and by The Next Web, which says it has not verified the intrusions; Microsoft and Unit 42, read directly.

People

A former OpenAI engineer charged over a recording device in an executive meeting room. Cross sectoral. [Reported]

• US prosecutors charged Abdulrahman Alfozan with using a device to intercept oral communications; OpenAI says it found the device on 30 September and referred it to the FBI (iTnews, from Reuters, 20 October).
• A detention filing alleges "customised, encrypted recording devices to secretly record highly sensitive meetings", and frequent contact with an unnamed member of the Saudi royal family.
• His lawyer says a planned Saudi trip was booked before the investigation. These are allegations.
• The charge itself came the week before the window; the filing is this week's.

Attribution: US prosecutors' filing as reported by Reuters through iTnews. One outlet read, so lighter.

The catalogue

Eight additions to the exploited catalogue in the window. Cross sectoral. [Confirmed]

• 20 October: four Plex Media Server flaws, CVE-2026-96652, 96654, 96655 and 96656; Siemens Software Center CVE-2026-110936; SiteGuard Server Edition CVE-2026-106931. 23 October: Arista VeloCloud Orchestrator CVE-2026-114625; Collabora Online's built in CODE server CVE-2026-110447. Ransomware use "Unknown" for all (KEV feed, version 2026.10.23, 1,756 entries).
• Since last week. SiteGuard and Collabora were in last week's Summary as exploited and not yet catalogued. Both now are.
• PHP2Shell is not listed. No Fortinet addition since CVE-2025-61624 on 16 October.

Attribution: CISA KEV, read directly.

Absences, counted rather than assumed

What was looked for and not found. Cross sectoral. [Confirmed]

• SEC EDGAR full text search for "Item 1.05", 19 to 25 October: one hit, a Vita Coco 10-Q of 22 October, and no 8-K. Quest Diagnostics' filing, in healthcare below, did not appear in it (SEC EDGAR).
• No fix found for Ivanti Connect Secure's third flaw, the privilege escalation in last week's Summary.
• No ACSC alert read: cyber.gov.au timed out, in a week ASD said it was affected by an outage. GreyNoise and Shadowserver were not checked this run.

Attribution: SEC EDGAR; the sources in each item above.

The week for a security leader

Energy and utilities

No in window incident at an energy or water operator met the bar. SafePay's advisory names critical infrastructure providers in Australia among its focus, and Microsoft lists energy among Storm-0940's targets. The PHP2Shell, SafePay and ConnectWise items in the cross sectoral section are this sector's exposure.

Financial services

State Bank of India finds 3,897 customer records on the dark web. [Confirmed]

• In an exchange filing on 21 October, SBI said alerts from 16 October led it to 3,897 customer records with personal information among about 375,000 records on the dark web, and that it expects no material impact on operations (The Hacker News; DTNext).
• How the data became available, and whether it came from SBI's systems, is not established in the filing as reported.
• Claimed. A group calling itself Siriusim claims 375,000 records and 2,722 PDFs on a criminal forum.
• India's Internet Freedom Foundation says it found no notice to customers (IFF).

Attribution: SBI's filing as reported by The Hacker News and DTNext, read through search; the filing was not read.

The ConnectWise and Canva items in the cross sectoral section touch every institution whose providers connect apps to its tenant or its customer records.

Healthcare

Quest Diagnostics confirms unauthorised access to its test ordering system. [Confirmed]

• Quest, in an SEC filing: "On October 21, 2026, Quest Diagnostics Incorporated learned of unauthorized access to its Quanum LSM system", the platform clients use to order tests and view results (BleepingComputer, 23 October).
• Quest believes some patient records were exposed; the data types are not yet determined. Quanum LSM is shut down, with paper requisitions in its place. Customer facing operations were not disrupted.
• Law enforcement notified. No group has claimed it, and it is not described as ransomware.

Attribution: Quest's filing as reported by BleepingComputer; the filing was not read.

The PHP2Shell and SafePay items in the cross sectoral section are this sector's other priorities.

Manufacturing and operational technology

No in window incident at a plant or in an OT environment met the bar. Two manufacturers' websites, Deere's and Cummins' vendor's, appear in the AI agent item in the cross sectoral section, without a confirmed loss of sensitive data, and APT28's campaign against a shipbuilder and an aircraft engine maker is under government and defence. The Arista and Fortinet items in the cross sectoral section are this sector's remote access exposure.

Technology and software

HPE patches eight maximum severity flaws in ArubaOS. [Reported]

• 72 vulnerabilities in ArubaOS, AOS-8 and AOS-10, across mobility conductors, access points, controllers and gateways; eight rated maximum severity, CVE-2026-79844 to CVE-2026-79851, authentication bypasses, command and SQL injection and buffer overflows leading to code execution (BleepingComputer, 23 October).
• HPE is "not aware of any public discussion or exploit code that targets the listed vulnerabilities as of the release date of this advisory."
• Until patched, restrict the management interfaces to a dedicated segment. Fixed versions are in HPE's advisory, hpesbnw05162en_us, not read.

Attribution: HPE as reported by BleepingComputer. One outlet read, so lighter.

Japan's Onamae Desktop Cloud: malicious programs on up to 26,739 servers. [Reported]

• GMO Internet confirmed unauthorised access on 20 October and announced on 21 October that malicious programs were installed on some users' servers, with up to 26,739 potentially affected. It is shutting servers down in sequence (BleepingComputer; ASCII.jp).
• GMO says it has not confirmed a leak of customers' personal data and holds no card details. Users are asked to change the IDs and passwords stored on their servers.
• Entry method not disclosed.

Attribution: GMO Internet as reported by ASCII.jp and BleepingComputer, read through search.

The PHP2Shell, eMudhra, LiteLLM and OVHcloud items in the cross sectoral section are this sector's priorities, and Canva's is the reminder that a connected tool reaches the customer records it connects to.

Government and defence

APT28 inside a naval shipbuilder in a NATO country. [Reported]

• Symantec attributes a campaign to APT28, which is linked to Russia's GRU and also tracked as Fancy Bear, BlueDelta and Forest Blizzard. Targets: a naval shipbuilder, a military aircraft engine maker and a law enforcement agency in two unnamed European NATO countries (The Record, 20 October).
• At the shipbuilder an employee opened a malicious Word document; the first confirmed infection was 24 July and activity ran to 24 September. Symantec's analyst says the campaign appears ongoing.
• New tooling: the HOOKEDGE backdoor, overlapping the older HEADLACE implant (Cybersecurity Dive; The Hacker News).
• No confidence level stated in the reporting read. Symantec's own report was not read.

Attribution: Symantec as reported by The Record, Cybersecurity Dive and The Hacker News.

The PHP2Shell outages, including cyber.gov.au and Australian government sites, and Unit 42's European government break-in are in the cross sectoral section, with SafePay and ConnectWise.

Retail and consumer

Capcom confirms a breach of more than 11.7 million member accounts. [Confirmed]

• Capcom detected unauthorised access to its CAPCOM ID membership system on 19 October and reported it to police and Japan's Personal Information Protection Commission on 21 October (BleepingComputer, 22 October).
• Exposed: email addresses with account creation dates, login counts and last login IP addresses. Dates of birth and other details for 43 accounts. Capcom says other members' passwords and card data were not exposed.
• Attackers changed passwords on 196 accounts to lock their owners out; Capcom reset them. Entry method not disclosed.

Attribution: Capcom as reported by BleepingComputer, read through search.

The AI agent item in the cross sectoral section is this sector's warning: the operator's agents went after ecommerce sites, Australia's among them. PHP2Shell and Canva are its other exposures.

Transport and logistics

No in window transport or logistics incident met the bar. Microsoft lists aerospace among Storm-0940's targets. The PHP2Shell, SafePay and edge items in the cross sectoral section are this sector's exposure.

Professional services

Bullhorn takes its staffing platform offline after a cyber incident. [Reported]

• Bullhorn: "We have detected a cyber incident on the Bullhorn Platform." It disabled access to contain it, says it has isolated the issue and eradicated the immediate threat with CrowdStrike, and has "not yet determined whether customer data was exfiltrated" (BleepingComputer, 19 October; Cyber Daily).
• Problems began the previous Friday, 16 October, as dated context. ATS and CRM, Back Office and Talent Platform were among the services down; no restoration date was given.
• Password resets for non SSO users and API password rotation are required. Bullhorn warns it will never ask for passwords or codes by email, phone or text.
• No group has claimed it.

Attribution: Bullhorn as reported by BleepingComputer and Cyber Daily.

The ConnectWise and Canva items in the cross sectoral section are this sector's most relevant: most firms rely on providers who connect into their systems. The former OpenAI engineer item applies to any firm that holds sensitive meetings.

Circulating but not carried

• The ABC wiping staff devices after spyware linked ads. Attributed to Cyber Daily citing ABC News on 23 October; neither page was found or read.
• "PHP2Shell affects 70 per cent of the internet." The figure measures how widely PHP is used, not how many sites are vulnerable.
• Palo Alto's 22.6 million potentially vulnerable PHP instances. A vendor exposure count; no methodology seen.
• Patchstack's "multiple active exploitation incidents from hosting partners", as Recorded Future reports it. Unnamed, and Patchstack's own report was not found.
• Mirai botnet operators joining PHP2Shell attacks. SecurityWeek refused the read.
• Two Australian organisations, an Anglican schools council and a technology firm, named in ransomware claims. Single source each, not read. Names held pending Adam's ruling.
• Seven Deadly Sins' 200 million Canva rows. One outlet, a claim, not verified.