{"uuid": "3d374920-f8d6-468c-b80f-3a1346f84fbc", "vulnerability_lookup_origin": "1a89b78e-f703-45f3-bb86-59eb712668bd", "author": "9f56dd64-161d-43a6-b9c3-555944290a09", "vulnerability": "CVE-2026-56164", "type": "seen", "source": "https://t.me/bhhub/1204", "content": "Weekly 8 AI & Cyber signals to act on (Jul 13\u201320, 2026)\n\n#AISecurity@bhhub\n\n\u2728 NadMesh industrializes attacks on exposed AI and MCP services\nXLab found a Go botnet combining Shodan harvesting, 20+ exploit paths, polymorphic builds, and persistence to steal cloud keys, Kubernetes tokens, and callable MCP tools. Sensors saw about 139 source IPs/day; the panel\u2019s 3,811-key claim conflicts with other counters, so treat scale cautiously.\n\n\u2728 Mandiant makes AI-assisted vulnerability management enforceable\nThe blueprint treats source code as untrusted agent input and puts deterministic policy before model judgment: isolated workloads, ZDR, short-lived repo-and-branch-bound identities, skill/MCP vetting, and runtime toxic-flow monitoring. Static scanning cannot stop post-deployment exfiltration; guard models cannot replace authorization or sandboxing.\n\n\u2728 Gold Eagle centralizes triage for AI-discovered vulnerabilities\nThe new US clearinghouse links government, AI labs, open-source maintainers, and critical infrastructure to deduplicate model-driven scanning and route fixes. Intake has begun, but no throughput, disclosure-SLA, or remediation metrics are public; performance remains unmeasured.\n\n#AppSec@bhhub\n\n\u2728 Update: a separate SharePoint RCE enters CISA KEV\nCVE-2026-58644 is an unauthenticated deserialization RCE scored 9.8. CISA added it to KEV on July 16 as actively exploited, automatable, and total impact. This is new evidence after the Jul 15 digest\u2019s separate CVE-2026-56164: patch affected on-prem 2016, 2019, and Subscription Edition builds.\n\n\u2728 Harden-Runner blocks undeclared CI egress on macOS and Windows\nVersion 2.20.0 extends egress allowlist enforcement to GitHub-hosted macOS and Windows runners handling Apple provisioning and signing keys. Teams can baseline in audit mode, then block undeclared endpoints so poisoned actions cannot exfiltrate secrets; the artifact is public, though efficacy claims are first-party.\n\n#RedTeam@bhhub\n\n\u2728 UAT-11795 hides resilient C2 behind installers, Telegram, and Polygon\nTalos found trojanized MobaXterm, WebEx, Zoom, DBeaver, and FACEIT installers delivering the new Starland RAT and WLDR memory implant. The actor can load CastleStealer or Remcos, stores fallback C2 in a Polygon contract, and sends beacons to Telegram\u2014two new implants plus durable hunting patterns.\n\n\u2728 Sandworm\u2019s UAC-0145 moves ClickFix infrastructure into smart contracts\nCERT-UA links at least 10 compromised sites, fake CAPTCHA commands, torrents, Signal lures, and counterfeit Android tools to UAC-0145. SMARTAXE selects content per visitor and retrieves its malicious domain from Ethereum before installing GHETTOVIBE; victim scale is undisclosed, but infrastructure can rotate without changing the page.\n\n#BlueTeam@bhhub\n\n\u2728 Microsoft maps two prevalent ACR Stealer chains\nOne ClickFix chain mounts WebDAV content, stages PowerShell/Python, persists by task, and can resolve C2 through blockchain RPC; the other uses MSHTA, JPEG steganography, and reflective execution. Both steal Chromium credentials, DPAPI tokens, and synced files; Microsoft reports increased activity but no victim denominator.", "creation_timestamp": "2026-07-20T10:00:04.912612Z"}