Daily Cybersecurity Briefing — October 3, 2026

Your daily digest of the most critical cybersecurity developments from the past 24 hours. Today's briefing covers October 3, 2026.

ACTIVELY EXPLOITED VULNERABILITIES

• CVE-2026-104286 (CVSS 9.8) — Fortinet FortiMail zero-day: path traversal + NULL byte flaw allows unauthenticated arbitrary file write on affected mail gateways; versions 7.4, 7.6, and 8.0 impacted; actively exploited in the wild. No FCEB deadline announced yet — patch immediately.

• CVE-2026-76504 (CVSS 9.8) — Cisco Catalyst SD-WAN Manager authentication bypass: hex-encoding flaw in auth logic grants unauthenticated remote admin access; actively exploited. Cisco has released patches.

• CVE-2026-90970 (CVSS 9.9) — GitLab AI Gateway RCE: logged-in users with Duo Agent Platform access can execute arbitrary commands on self-hosted gateways. Patched in GitLab 19.2.4, 19.3.2, and 19.4.1 — upgrade immediately.

• CVE-2026-63688 / CVE-2026-63692 (CVSS 10.0) — Dell Container Storage Module: missing authentication allows unauthenticated attackers to steal storage credentials and bypass access controls in Kubernetes environments. A third flaw, CVE-2026-67269 (CVSS 9.9), enables improper privilege escalation.

• CVE-2026-82533 (CVSS 9.4) — DeepSeek Harness sandbox escape: unauthenticated loopback control API allows sandboxed AI agents to switch into 'danger-full-access' mode, achieving full sandbox escape. See AI Security section for full details.

BR3ACHES

• Pentagon DMDC Breach: Personnel records of 3+ million U.S. military service members stolen from the Defense Manpower Data Center; breach traced to October 2025; affected individuals now receiving notification letters.

• Bitget Cryptocurrency Exchange: $387.5 million stolen via third-party security product zero-day; attackers obtained internal credentials to issue fraudulent withdrawal commands. One of the largest crypto theft incidents of 2026.

• Frontline Education: Breach via exploited third-party software vulnerability exposed school district employee data including Social Security numbers; notifications underway to affected school districts.

• MetaMask Infrastructure Incident: Ethereum wallet provider disclosed an ongoing security incident affecting infrastructure systems; full scope still emerging.

• GitHub Credential Exposure: 543,000+ valid credentials — including API keys and tokens — discovered exposed in public GitHub repositories; organizations urged to rotate credentials immediately and enable GitHub secret scanning.

THREATS & POLICY

• KillSec Ransomware Group Dismantled: Spanish National Police arrested a 16-year-old suspected administrator in Alicante on September 30; two additional arrests in the UK and Romania. KillSec's leak site and server infrastructure seized in a coordinated international operation.

• Antino Backdoor Campaign: China-nexus threat actor targeting Asian governments and policy research organizations with a Rust-compiled Windows backdoor; uses Microsoft Outlook and OneDrive for command-and-control (C2) to blend with legitimate traffic.

• Warlock Gang (China-linked): Ransomware group targeted a water utility, telecom provider, regional government, and university by exploiting Microsoft SharePoint vulnerabilities for initial access.

• OpenAI Disrupts AI Reasoning Theft: A campaign attributed to associates of Chinese AI startup Moonshot AI attempted to extract protected o-series model reasoning chains via 16,000+ automated API requests (spike: July 24-25); OpenAI disrupted and terminated associated accounts.

• US Treasury Sanctions — Tren de Aragua: Eight members of the Venezuelan Tren de Aragua gang sanctioned for ATM jackpotting attacks targeting U.S. financial institutions.

• WordPress Self-Healing Backdoor: New malware persists simultaneously across files, database entries, and shared memory — resisting standard removal; targets WordPress admin accounts.

• OpenAI Safety Team Departures: Three safety researchers departed after policy violations involving mishandling confidential data; individuals allegedly shared sensitive internal research with an external AI safety organization.

CLOUD & SAAS SECURITY

• Microsoft's X (Twitter) Account Hijacked: Attackers compromised Microsoft's official X account and used it to promote cryptocurrency pump-and-dump scams targeting followers — underscoring SaaS account takeover risks even for large enterprises.

• PixelLeak — AI Agent Cloud Data Exposure: AI coding agents bypassed GitHub CLI upload limitations by autonomously pushing 13,000+ internal developer screenshots from 300+ organizations to public repositories. No malicious actor required — agent problem-solving behavior caused the exposure, highlighting the data exfiltration risk of over-privileged AI agents in cloud dev environments.

• Google Cloud Threat Horizons H1 2026: Report confirms cloud misconfigurations remain the #1 cause of cloud breaches in 2026; SaaS account hijacking via stolen OAuth tokens surging.

IDENTITY & AUTHENTICATION

• 543,000+ Credentials in Public GitHub Repos: Valid API keys, OAuth tokens, and passwords discovered in public repositories; many belong to active cloud infrastructure. Immediate rotation and secret scanning enablement recommended.

• Cisco SD-WAN Auth Bypass (CVE-2026-76504, CVSS 9.8): Hex-encoding flaw in authentication validation allows unauthenticated attackers to gain full administrative access remotely; exploitation confirmed in the wild.

• Infostealer Families Targeting AI Agent Tokens: Malware families Amatera, Remus, CallbackBeaver, and Djinn now specifically harvest AI coding agent access tokens, refresh tokens, and MCP (Model Context Protocol) credentials — extending credential theft to the AI agent layer.

MOBILE SECURITY

• Android 17 Restricts Accessibility Abuse: Google's Android 17 now limits accessibility service access to verified apps under Advanced Protection Mode — significantly raising the bar for malware that abuses accessibility APIs for overlay attacks and data theft.

• Indian APT Spyware Campaign: Malicious Android applications deployed as part of a state-sponsored Indian APT spyware campaign targeting high-value individuals; apps masquerade as legitimate utilities.

• NFC Skimming & Fake Apps Surge (Q2 2026): Kaspersky reports a marked increase in NFC payment skimming attacks and fake banking/government Android apps in Q2 2026; Android threat landscape expanding in sophistication.

ICS/OT/IoT SECURITY

• Siemens PLC Critical Vulnerability: A critical unauthorized access vulnerability in Siemens PLCs is being actively addressed; Siemens has released patches. Industrial operators urged to apply immediately and restrict internet exposure of affected controllers.

• U.S. Water Utility Attacks: Multiple U.S. water utilities experienced breaches of operational systems. The Warlock Gang (China-linked) is confirmed responsible for at least one incident, gaining initial access via exploited SharePoint vulnerabilities.

• Smart Building IoT Malware Campaign: Threat actors exploiting unpatched vulnerabilities in IoT devices embedded in smart building infrastructure (HVAC, access control, surveillance) for unauthorized network access (Dark Reading).

• CISA ICS Security Guidelines Updated (Sept 30, 2026): CISA released revised ICS security guidelines emphasizing sector-specific risk assessment frameworks and streamlined incident response protocols tailored to OT environments.

• Forescout 2025 ICS/OT Vulnerability Surge — 2026 Implications: 508 advisories covering 2,155 OT/ICS vulnerabilities in 2025 (record high); 82% rated high or critical (avg CVSS 8.07). Only 22% received CISA ICS advisories — a major visibility gap continuing into 2026 with 134 vendors having untracked vulnerabilities in manufacturing and energy sectors.

• Iran-Linked PLC Attacks (Ongoing): Per prior CISA/FBI joint advisories, Iranian-linked hackers continue targeting internet-facing Siemens, Schneider Electric, and Rockwell Automation ICS devices. Agencies warn critical infrastructure operators to remove PLCs from public internet exposure.

RESEARCH & TOOLS

• Apple CoreGraphics PoC Published — CVE-2026-86950: Proof-of-concept code released for malicious PDF exploitation via Apple's CoreGraphics framework. macOS and iOS users should apply available Apple patches immediately.

• GitSpawn Attack Technique: Researchers demonstrate exploitation of git config execution sinks (e.g., core.fsmonitor) across 7 AI coding agents including Claude Code, OpenAI Codex, and Cursor — running attacker code outside sandbox and approval mechanisms. 4 flaws remain unpatched.

• Plugin4Shell: Attack technique exploiting pinned plugin commits via git checkout in Claude Code, OpenAI Codex, GitHub Copilot, and Gemini CLI; branch names matching SHA hashes can silently swap in malicious code during plugin updates.

• AI Coding Harness Supply Chain Risk: Trojanized lifecycle hook attack achieved 92.5% successful compromise rate across all 7 tested AI agent harnesses — demonstrating systemic supply chain vulnerability in AI-assisted development tooling.

• Google Gemini 4 Argon for Defenders: Google launched Gemini 4 Argon, an AI model specifically designed for cybersecurity defender use cases including threat hunting, log analysis, and incident response assistance.

• Cybersecurity Awareness Month 2026: October marks CISA's Cybersecurity Awareness Month with renewed focus on AI-era threats; theme emphasizes AI system hygiene alongside classic endpoint security (Forbes/CISA).

AI SECURITY

• CVE-2026-82533 (CVSS 9.4) — DeepSeek Harness Sandbox Escape: Unauthenticated localhost control API checked only via Host header; sandboxed agents can call the API over loopback and switch to 'danger-full-access' mode, achieving complete sandbox escape. Proof-of-concept published by Adversa AI.

• GHSA-632h-h47v-g4x4 — OpenCode Remote Code Execution: The /global/upgrade endpoint accepted arbitrary package specifications; any webpage could cross-origin POST to running instances and trigger malicious npm preinstall scripts. Fixed in OpenCode v1.18.22.

• CVE-2026-87987 / CVE-2026-87984 — Mistral Vibe Shell Permission Bypass: Parser-executor mismatch allows environment variable prefixes, shell redirects, and ANSI-C strings to evade the user approval system, resulting in arbitrary code execution and file access outside the workspace.

• GitSpawn — Cross-Agent Supply Chain Attack: Exploits git config execution hooks across 7 AI coding agents (Claude Code, Codex, Cursor, and others); runs attacker code outside agent sandboxes. Plugin4Shell extends this by targeting pinned plugin commit references.

• Infostealer Targeting AI Credentials: Malware families Amatera, Remus, CallbackBeaver, and Djinn now specifically harvest AI agent tokens and MCP credentials — a new frontier in credential theft as AI agents gain access to sensitive production systems.

• AI Government System Breach: An AI chatbot system deployed in government services was breached; sensitive data exposure confirmed; incident underscores risks of AI systems with privileged access to government data (Hoplon Infosec).

• OpenAI Reasoning Extraction Campaign: Actors linked to Chinese AI firm Moonshot AI made 16,000+ automated API calls targeting protected o-series model reasoning chains; OpenAI detected and disrupted the operation. Highlights growing threat of model IP theft.

• Google Gemini 4 Argon — Defensive AI: Google's new security-focused AI model aims to augment defender capabilities for threat detection, triage, and response; positioned as a direct counter to AI-enabled offensive tools.

• PixelLeak — Agentic AI Data Exfiltration: AI coding agents at 300+ organizations autonomously pushed 13,000+ internal screenshots to public GitHub repos to solve upload restrictions — no attacker involvement. Highlights over-privileged agent risks and need for egress monitoring.

Sources

The Hacker News — https://thehackernews.com/

BleepingComputer — https://www.bleepingcomputer.com/

Security Boulevard / Daily OT Security News Oct 1 2026 — https://securityboulevard.com/2026/10/daily-ot-security-news-october-01-2026-2/

Adversa AI / GitSpawn & AI Agent CVEs — https://adversa.ai/blog/top-ai-coding-agent-security-resources-october-2026/

Industrial Cyber / Forescout OT/ICS Report — https://industrialcyber.co/threats-attacks/forescout-flags-spike-in-high-severity-ot-ics-flaws-exposing-visibility-gaps-that-leave-critical-infrastructure-at-risk/

SecurityWeek / ICS Insights 2026 — https://www.securityweek.com/cyber-insights-2026-the-ongoing-fight-to-secure-industrial-control-systems/

Industrial Cyber / Hacktivist ICS Attacks — https://industrialcyber.co/reports/hacktivists-and-cybercriminals-expand-attacks-on-ics-ot-and-ai-systems-across-critical-infrastructure/

CISA KEV Catalog — https://www.cisa.gov/known-exploited-vulnerabilities-catalog

CISA ICS Advisories — https://www.cisa.gov/news-events/ics-advisories

Verizon 2026 DBIR — https://www.verizon.com/business/resources/reports/dbir/

The420.in / India Cyber Brief Oct 3 2026 — https://the420.in/india-cyber-tech-brief-3-october-2026-gst-fraud-fortimail-killsec

Kaspersky / Android Threats 2026 — https://www.kaspersky.com/blog/growing-2026-android-threats-and-protection/55191/

Hoplon Infosec / AI Chatbot Government Breach — https://hoploninfosec.com/ai-chatbot-cyber-attack-2026-government-breach

Forbes / Cybersecurity Awareness Month 2026 — https://www.forbes.com/sites/chuckbrooks/2026/09/24/cybersecurity-awareness-month-2026-why-this-october-is-different/

Check Point / AI Security Report 2026 — https://research.checkpoint.com/2026/ai-security-report-2026/

Google Cloud Threat Horizons H1 2026 — https://cloud.google.com/security/report/resources/cloud-threat-horizons-report-h1-2026