Daily Cybersecurity News – October 22, 2026

What happened

Amazon Web Services detected exploitation attempts against the PHP2Shell pair of critical PHP vulnerabilities on its honeypots starting October 17, four days before public disclosure and patches on October 21.

Proof-of-concept exploits appeared around disclosure time, and AWS reported the flaws are now being broadly exploited in the wild. Attempts came from anonymized IPs linked to commercial proxy or VPN networks, with multiple actors active in the hours before the official release.

Who is affected

Any internet-facing systems or applications running unpatched versions of PHP vulnerable to the PHP2Shell flaws. PHP powers a large share of the web, so the blast radius includes countless websites, CMS platforms, and custom apps.

Administrators who delayed patching after the Wednesday release remain exposed while active scanning and exploitation continue.

Why it matters

Pre-disclosure exploitation shows sophisticated actors obtained details during embargo, shrinking the window between patch and weaponization. AI tools accelerate reverse-engineering of fixes into working exploits, lowering the barrier for less skilled attackers.

Operators face immediate risk of remote code execution on unpatched PHP instances, with potential for widespread compromise of web infrastructure. Coordinated disclosure processes need to adapt to faster analysis cycles.

How it could have been prevented

Apply the official PHP patches for the two critical vulnerabilities immediately on all affected systems. Prioritize internet-facing hosts and use honeypots or logging to detect scanning.

Monitor for anomalous traffic from anonymization networks, isolate vulnerable instances where possible, and verify no successful exploitation occurred before patching. Keep PHP and dependent applications updated going forward.

What happened

Microsoft attributed a supply-chain attack on ConnectWise Cloud Backup customers' Microsoft 365 environments to the China-state aligned group Storm-0940. The group used the open-source agentic AI penetration testing framework ARTEX to automate reconnaissance, password spraying, and post-authentication actions.

Activity began as early as September 2026. A preceding password-spray campaign hit 327 accounts across eight organizations with over 151,000 failed logons peaking at 2,900 per minute. Storm-0940 then leveraged the same infrastructure and AI agents to compromise ConnectWise customers, reaching more than 2,250 tenants. In some cases agents progressed to Microsoft Graph operations, credential changes, and privilege escalations.

Who is affected

ConnectWise Cloud Backup customers whose Microsoft 365 environments were accessible via the compromised third-party integration. Targets spanned aerospace, defense, automotive, energy, finance, government, manufacturing, tech, and other sectors globally.

Hundreds of organizations were scanned or attacked; Microsoft observed follow-on activity in multiple customer environments.

Why it matters

AI agents dramatically accelerate attack tempo and scale, allowing nation-state actors to spray credentials and abuse high-privilege OAuth apps across thousands of tenants with minimal human effort. A single compromised SaaS backup provider can cascade access far beyond its intended scope.

Builders and operators of third-party integrations must assume supply-chain abuse will be automated and treat elevated app permissions as high-risk. Multiple actors are already adopting ARTEX.

How it could have been prevented

Review and minimize OAuth application permissions for ConnectWise and similar third-party apps in Microsoft 365. Rotate credentials, enable conditional access and phishing-resistant MFA, and monitor for anomalous Graph API and directory activity.

Hunt for signs of password spraying and unexpected app credential modifications. Isolate or revoke high-privilege third-party integrations until verified secure, and apply least-privilege principles to all connected services.

What happened

GitLab released patches on October 21 for CVE-2026-110264, a CVSS 10.0 unauthenticated remote code execution vulnerability caused by an integer overflow in the Onigmo regular-expression compiler. It affects every self-managed GitLab CE/EE version from 7.0 onward.

The flaw can allow an unauthenticated attacker to execute arbitrary code under certain conditions via a crafted regular expression. No active exploitation has been observed and it is not in CISA KEV. A prior related fix on October 12 covered only 19.2 and later, leaving older and some recently patched instances still vulnerable until the new updates.

Who is affected

All self-managed GitLab instances running versions from 7.0 before the fixed releases: 18.11.13, 19.0.10, 19.1.9, 19.2.9, 19.3.6, 19.4.3, or 19.5.1. GitLab.com and properly updated Dedicated tenants are already protected.

The vulnerability reaches back more than a decade because GitLab ships and patches its own Onigmo engine.

Why it matters

Unauthenticated RCE on a widely used DevOps platform gives attackers full control of source code, CI/CD pipelines, secrets, and potentially the underlying host. The 12-year span means many long-lived or poorly maintained instances are exposed.

Even admins who applied the October 12 patches must upgrade again. Lack of public preconditions means no configuration-based exemption; patching is the only reliable fix.

How it could have been prevented

Upgrade immediately to the fixed version for your release branch: 19.5.1, 19.4.3, 19.3.6, 19.2.9, 19.1.9, 19.0.10, or 18.11.13. GitLab backported the fix beyond its normal support window because of severity.

Use the published Sigma detection rules for Onigmo worker crashes and Org mode render error bursts to hunt for exploitation attempts. There is no workaround; apply the patch and restart services.

What happened

North Korean Lazarus Group operators used the standardized ML-KEM post-quantum key encapsulation mechanism to protect malware payload delivery in a campaign Kaspersky calls FlashHole. Attacks against South Korean organizations began around July 2025 and combine known and new tools.

The chain starts with phishing to crafted GitHub pages that load external JavaScript, abuse the local INISAFE CrossWeb EX security tool, and run shellcode via fsquirt.exe to install the TechHub downloader. LPEClient then obtains ML-KEM-1024 public keys from C2, derives shared secrets, and decrypts subsequent payloads. Additional malware includes ClientHub, the modular Cubridge backdoor, and Chromium credential stealers.

Who is affected

Primarily South Korean organizations in software, construction, and wired communication equipment manufacturing. Kaspersky confirmed at least three victims and believes more were hit.

Any organization in the region or supply chain that could be reached via the phishing and GitHub-based delivery remains at risk while the campaign continues.

Why it matters

Adoption of post-quantum cryptography by a sophisticated APT raises the bar for defenders who still rely on classical crypto analysis for traffic inspection or payload recovery. It also signals future-proofing of C2 channels against quantum-capable adversaries.

The blend of living-off-the-land binaries, legitimate Korean security software abuse, private GitHub repos, and modular backdoors creates a resilient, multi-stage infection that is harder to detect and attribute in real time.

How it could have been prevented

Block or tightly control access to untrusted GitHub pages and external script loads. Monitor for anomalous use of fsquirt.exe, INISAFE processes, and unexpected Dropbox or private-repo network connections.

Deploy behavioral detection for LPEClient, TechHub, ClientHub, and Cubridge indicators. Enforce application allow-listing, isolate high-risk users, and rotate credentials on any suspected compromise. Keep endpoint and network telemetry focused on in-memory execution chains.

What happened

PaaS provider Railway disclosed anomalous activity on systems storing customer environment variables. An unauthorized party may have accessed encrypted environment variables and, in some cases, obtained decrypted contents containing secrets such as API keys and credentials.

Impact is limited to a subset of users. The company links the incident to a broader campaign against platform providers aimed at reaching AI inference services. Attackers reportedly drained cryptocurrency wallets whose private keys were stored as Railway environment variables, with theft beginning hours before public disclosure. Railway rotated its own bucket credentials and invalidated API tokens.

Who is affected

A subset of Railway customers who stored sensitive environment variables (API keys, database credentials, registry tokens, crypto keys, etc.). Exact numbers were not published.

Downstream AI inference providers whose keys were stored on Railway may also have been exposed; Railway notified them to revoke keys.

Why it matters

Environment variables are a common secret store on PaaS platforms. Compromise of the platform's secret storage immediately endangers every customer workload that relies on those values, enabling follow-on abuse including crypto theft and LLMjacking (running attacker LLMs on victim cloud bills).

Platform customers must treat hosted secret stores as high-value targets and maintain independent rotation and monitoring practices.

How it could have been prevented

Immediately rotate all environment variables and secrets stored in Railway, including API keys, private registry credentials, and any other sensitive values. Review service logs for unusual activity.

Railway has already rotated its internal credentials; customers should re-add new values and verify builds. Monitor for unauthorized use of any exposed keys and enable alerts on anomalous cloud or AI API usage.

What happened

Hackers breached the infrastructure of Onamae Desktop Cloud (by GMO Internet) and installed malicious programs on up to 26,739 customer VPS servers. The company confirmed unauthorized access on October 20 and began progressively shutting down affected servers from 04:00 local time on October 21.

Data stored on the servers may have been exposed, though no confirmed misuse or leak has been reported. Payment card data sits on separate infrastructure and is unaffected. The service, popular for continuous FX/MetaTrader automated trading, remains interrupted with no restart ETA.

Who is affected

Up to approximately 27,000 Onamae Desktop Cloud customers in Japan whose VPS instances received malware. All Desktop Cloud plans are impacted by the shutdown.

Users who stored passwords, trading credentials, or other secrets on the servers face elevated risk of compromise and follow-on phishing.

Why it matters

A provider-level compromise that reaches tens of thousands of customer servers at once demonstrates the concentrated risk of VPS and desktop-cloud platforms, especially those used for always-on financial trading bots. Malware installation enables data theft, further lateral movement, or ransomware.

Customers lose availability immediately and must assume secrets on disk are burned. Attackers may weaponize the incident for targeted phishing.

How it could have been prevented

Rotate all passwords, API keys, trading platform credentials, and other secrets that were present on the affected servers as soon as possible. Monitor for phishing emails or calls impersonating Onamae/GMO.

Do not provide credentials if contacted; the company states it will never ask for passwords or card details via email or phone. Wait for official restoration guidance and rebuild instances from clean images once service returns.

What happened

Security firm Tracebit released research on 'context bombs' - crafted hidden prompts automatically ingested by attacking AI agents when they crawl a target. The deceptive content misdirects agents toward decoys, endless loops, or fake high-value targets, causing them to waste effort or abandon real objectives.

In tests the technique stopped all attacks from several models (GPT 5.6 Cyber, Kimi K3, Qwen 3.8 Max, DeepSeek V4 Pro). It cut Claude Opus 5 success by 91% and GPT-6 Astra by 54%, while nearly quadrupling or multiplying the compute cost of successful attacks up to 23-fold. The strongest single bomb was a fake penetration-test report that labeled real paths as canaries and decoys as already-exploited prizes.

Who is affected

Organizations concerned about autonomous AI agents performing reconnaissance or exploitation against their public or internal assets. The defensive technique is relevant to any defender facing LLM-driven offensive tooling.

It applies broadly as more threat actors adopt agentic frameworks for automated attacks.

Why it matters

As AI agents lower the skill and time required for complex attacks, low-cost defensive deception that inflates attacker compute bills and reduces success rates becomes a practical force multiplier. Context bombs turn the agent's own context window against it without requiring detection of every probe.

Builders of agentic systems and defenders can both learn from the failure modes: agents remain highly suggestible to planted narrative and can be trapped by recursive or high-complexity decoys.