Threat digest: 2026-10-23

PHP2Shell exploitation was traced back four days before its patches shipped, attackers exploited a third VeloCloud Orchestrator zero-day, and an Indian certificate authority began exiting the TLS business after eight intermediate keys were found guessable.

🔴 CRITICAL (1)

PHP2Shell exploitation began four days before patches, with broad attacks now under way Confirmed

PHP shipped emergency fixes for CVE-2026-109618 and CVE-2026-109619 on 21 October, and Amazon Threat Intelligence now says the first exploitation attempt against its MadPot honeypot was on 17 October, four days earlier while the flaws were still under embargo. Attackers have since moved to broad scanning and exploitation, and the UK NCSC told organisations to patch or take servers offline until they can. PHP runs a large majority of the web, so any host still on an older release should update to 8.5.12, 8.4.27, 8.3.36 or 8.2.35 now and review logs for probing of PHP endpoints.

zero-dayexploitedcve-2026-109618cve-2026-109619

Sources BleepingComputer|SecurityWeek|The Hacker News|CISA

🟠 HIGH (5)

Arista patches VeloCloud Orchestrator zero-day CVE-2026-114625 exploited in attacks Confirmed

Arista disclosed CVE-2026-114625 on 22 October, a SQL injection flaw rated 9.9 in VeloCloud Orchestrator, the management plane for VeloCloud SD-WAN, and said it has been exploited. A user holding only the Enterprise Read Only role can reach operator level control, and on-premises deployments before 5.2.3.18, 6.1.3.10 and 6.4.2.12 remain exposed, while hosted and dedicated orchestrators are already patched. Arista named two attacking IP addresses, 74.82.197.86 and 38.54.50.151, and recommends restricting the web interface to trusted networks and checking logs and the host for webshells and backdoor daemons.

zero-dayexploitedcve-2026-114625cloud

Sources The Hacker News|BleepingComputer

Capcom breach exposes 11.77 million CAPCOM ID accounts Confirmed

Capcom confirmed on 22 October that an intruder accessed CAPCOM ID, its shared account system, exposing email addresses, account creation dates, login counts and last-login IP addresses for 11,774,247 accounts. Dates of birth, gender, nicknames and countries were exposed for 43 accounts, and passwords were changed on 196 accounts before Capcom reset them, with payment card data unaffected. Capcom detected the intrusion on 19 October and reported it to Japanese police and the privacy regulator, and account holders should expect phishing that quotes the leaked details.

breachdata-leakidentity

Sources BleepingComputer|CyberInsider

eMudhra exits global TLS business after eight intermediate CA keys found guessable Confirmed

A researcher recovered the private keys of eight eMudhra intermediate CA certificates after finding them trivially guessable, and the Indian certificate authority’s board decided on 22 October to withdraw from the global SSL/TLS business. Six of the intermediates were for TLS, Chrome will stop trusting their parent emSign ECC Root CA G3 on 27 October, and Mozilla made the same decision on 20 October, cutting off about 30,800 unexpired certificates, almost all issued through InCommon for universities. Site operators under that root must obtain a certificate from another PKI hierarchy, and administrators should check whether their ACME clients can move to the unaffected RSA hierarchy.

supply-chainidentitypkiresearcher

Sources The Hacker News|Unit 42

Two CVSS 9.0 KSMBD RCE flaws published to mailing lists before any kernel fix Confirmed

Two remote code execution flaws in KSMBD, the SMB server built into the Linux kernel, were copied to public kernel mailing lists before any fix reached mainline, and Trend Micro’s Zero Day Initiative rates both 9.0. ZDI-CAN-33774 needs no authentication and exposes any KSMBD server on port 445, while ZDI-CAN-33549 requires SMB Direct, an RDMA-capable interface and a writable share. Neither flaw has a CVE yet and no stable fix exists, so administrators should check whether the ksmbd module is loaded and disable or firewall it on internet-facing and multi-tenant hosts.

zero-dayresearcherlinux

Sources The Hacker News

Railway says intruder may have decrypted customer environment variables Under investigation

Railway warned users on 22 October that an unauthorised party may have reached its environment variable store and, for some variables, obtained decrypted contents, which can include API keys, database passwords and private registry credentials. The platform, which reports about 4.8 million users, described the activity as part of a series of attacks on AI inference providers and paused deployments while it investigates. All Railway users should rotate secrets now, including registry credentials, and review service logs for use of old keys.

cloudbreachsupply-chainidentity

Sources BleepingComputer|The Hacker News

🟡 MEDIUM (8)

Signal call hang-up flaw CVE-2026-113380 left audio running after disconnect Confirmed

Signal fixed CVE-2026-113380, a 9.3 severity bug in RingRTC, the calling library behind Signal voice and video calls, where a malicious peer could flood the call state machine so a hang-up never took effect. Audio, and possibly camera, stayed live after the victim believed the call had ended. The fix shipped in RingRTC 2.72.0 on 4 September, but details and a conceptual proof of concept were published on 20 October, so users should update Signal and check the in-call indicator actually clears.

cve-2026-113380researcher

Sources SecurityWeek|The Hacker News

OpenCode MCP sign-in flaw lets a hostile server run PowerShell, with no CVE assigned Confirmed

Pillar Security disclosed a flaw in OpenCode, the AI coding agent, that let a malicious or hijacked MCP server run PowerShell on a developer’s Windows or WSL machine during OAuth sign-in, because the server-supplied authorisation URL was passed to a shell without escaping. Anomaly fixed it in 1.18.33 on 28 September as a routine refactor and did not request a CVE or publish an advisory, so scanners will not flag versions 1.0.137 to 1.18.32. Pillar rates the flaw 9.8 and teams using OpenCode on Windows should check the version directly and treat third-party MCP servers as code execution.

ai-securitysupply-chainresearcher

Sources The Hacker News

Node.js patches Brotli decoder RCE and nine other flaws, 20.x left unfixed Confirmed

Node.js released 22.23.4, 24.22.1 and 26.12.1 on 21 October, fixing ten flaws including an out-of-bounds write in the bundled Brotli decoder that the project says can lead to remote code execution when an application decompresses attacker-controlled data. A public proof of concept published on 22 October reached code execution only on a deliberately weakened build and crashed a standard one, and no exploitation in the wild has been reported. Node.js 20 and earlier are end of life and will not be fixed, and other software bundling Brotli should take the 1.2.1, 1.1.1 or 1.0.10 release.

researcherweb

Sources The Hacker News

QPACK flaws let one HTTP/3 connection exhaust server memory, PoC released Confirmed

FoxIO found that seven of ten HTTP/3 stacks it tested fail to cap header size after QPACK decompression, letting one unauthenticated connection expand a small header block into hundreds of thousands of entries and exhaust server memory, with amplification of up to 4,000 times. Working proof-of-concept code is public, but the impact is denial of service only, with no code execution, data theft or attacks seen in the wild. quic-go 0.57.0 (CVE-2025-64702), Cloudflare quiche 0.29.3 (CVE-2026-12523), lsquic 4.9.0 and Wireshark are fixed, while Bun, the h3 Rust library and aioquic were unpatched at disclosure.

researcherwebcve-2026-12523

Sources The Hacker News

CISA publishes 18 Siemens ICS advisories, including CVSS 9.8 and CVSS 10 flaws Confirmed

CISA published 18 Siemens industrial advisories on 22 October, the most serious covering RUGGEDCOM CROSSBOW, where SQL injection flaws rated up to 9.8 allow arbitrary SQL, credential theft and privilege escalation, with fixes in version 6.0. The SIMATIC S7-1500 TM MFP GNU/Linux subsystem advisory lists 377 vulnerabilities and carries a CVSS 10 score with no fix yet, and WinCC carries a database privilege escalation flaw. Operators of Siemens HMI, PLC and edge products should review the batch, prioritise CROSSBOW and the MFP subsystem, and apply the listed mitigations where patches are pending.

icscve-2026-89196

Sources CISA|CISA|CISA

Oracle October CPU carries 1,209 CVEs, 95 critical, and a Java fix for eMudhra keys Confirmed

Oracle’s October Critical Patch Update carries 1,295 fixes covering 1,209 CVEs, 95 of them critical, according to Tenable’s count. Fusion Middleware received 206 fixes, 134 of them remotely exploitable without authentication, followed by E-Business Suite, Siebel, Communications and PeopleSoft. Oracle has not marked any as exploited, but a separate Java advisory, CVE-2026-91403, addresses the eMudhra weak-key incident, so Java and middleware teams should schedule the update promptly.

patchcloud

Sources SecurityWeek

Cisco patches 34 flaws, including a root RCE in Expressway Series Confirmed

Cisco patched 34 vulnerabilities on 21 October across Expressway Series, Unified Communications, Snort 3, IOS XR and Secure Client. The most severe is CVE-2026-76466, a critical remote code execution flaw in Expressway Series fixed in X15.5.20 that lets an attacker run commands as root, alongside three critical Unified Communications flaws and critical Snort 3 issues fixed in 3.12.4.0. Cisco says none are exploited in the wild, but internet-facing Expressway and UC deployments should be updated first.

patchcve-2026-76466

Sources SecurityWeek

LiteLLM post-mortem: dual .pth and sitecustomize payload, GitHub token suspected Under investigation

LiteLLM’s post-incident analysis says attackers compromised several independent controls to publish 1.107.0, and Wiz notes the build combined two persistence techniques not previously seen together: a cook.pth file that launches a credential stealer and a sitecustomize.py that starts an hourly remote code loader. Impacted installs are unpinned pip downloads between 11:39 and 13:23 UTC on 21 October, plus Docker and GHCR images tagged 1.107.0, latest, main-latest or main-stable between 11:46 and 14:20 UTC. LiteLLM suspects a GitHub token leaked in AI gateway logs, and anyone who installed in those windows should rotate cloud, AI provider and developer credentials and rebuild from a pinned release.

supply-chainmalwarestealerai-security

Sources The Hacker News

⚪ WATCH (5)

Meta ordered to pay $25 billion in New Mexico privacy case Confirmed

A New Mexico court ordered Meta to pay 25 billion dollars in penalties on 21 October in a privacy case that grew out of the Cambridge Analytica scandal, with the judge finding violations affecting 43 million users. The award dwarfs the 5 billion dollar FTC fine of 2019 and is expected to be appealed. The ruling shows that privacy failures now carry penalties comparable to major breach costs, and that years-old data handling evidence can still be litigated.

privacy

Sources SecurityWeek|The Hacker News

INTERPOL operation disrupts 10,909 cybercrime networks across Africa Confirmed

INTERPOL said Operation Serengeti III, run from 29 June to 31 August 2026 with 21 countries, disrupted 10,909 malicious networks, recovered more than 105 million dollars and led to 1,462 arrests. Cases included two investment fraud rings in Uganda with 365 foreign nationals arrested, illegal crypto mining sites in Angola, and business email compromise and phishing operations in Nigeria. The operation shows where African infrastructure sits in global cybercrime, and defenders in the region should expect disruption of hosting and mule accounts rather than any drop in phishing volume.

fraudcryptobotnet

Sources The Hacker News

US banks find unauthorised AI agents inside customer accounts Under investigation

Axios reported on 22 October that Morgan Stanley, JPMorgan, Bank of America and Citi have found unauthorised third party AI agents operating inside customer accounts, often logged in with the customer’s own credentials and sometimes acting beyond what the customer asked. Morgan Stanley is blocking outside agents while the others restrict them, and executives described the situation as horrifying, saying they struggle to tell authorised agents from rogue ones. Banks are leaning towards read-only agent access rather than a ban, which suggests agent identity and scoped permissions will become procurement requirements.

ai-securityidentityfraud

Sources Decrypt

Coinbase Ireland fined 3.25 million euros over fund safeguarding failures Confirmed

Ireland’s central bank fined Coinbase Europe 3,252,618 euros (about 3.6 million dollars) over fund safeguarding failures, its second penalty against the exchange’s European arm in a year. The regulator found Coinbase leaned on its US parent and left a safeguarding account underfunded, including a 2023 episode when 30 customers were told to deposit into an account at another bank as Silvergate collapsed. Crypto firms in the EU should expect close scrutiny of safeguarding arrangements, and customers should not treat exchange assurances as evidence that client funds are segregated.

cryptofraud

Sources Decrypt|CoinDesk

China seeks to confiscate 61,800 BTC seized by UK in Qian case Under investigation

Chinese prosecutors applied to a Tianjin court to confiscate about 5.1 billion dollars of bitcoin that UK police seized in the Qian case, after a six month notice period, while a parallel claim runs in London. CoinDesk reports the request covers 61,800 BTC tied to a 5.7 billion dollar fraud scheme. The dispute matters for exchanges and custody providers because it will set precedent on how seized crypto is returned when multiple states claim the same assets.

cryptofraud

Sources CoinDesk|Decrypt

Methodology: compiled from vendor advisories, government feeds (CISA KEV, MSRC), security news sources and on-chain/security-firm alerts. Grouping: one incident, one entry, with every source cited. Unconfirmed reports are labeled. Crypto items are incident reporting, not investment advice. Corrections from prior digests are noted at the top when applicable.