An unpatched Microsoft Defender zero-day, dubbed “RoguePlanet” (CVE-2026-50656), is being weaponized with a public exploit that hands attackers SYSTEM-level access on fully patched Windows 10 and 11 machines. No fix has shipped yet, which makes this a containment-and-detection problem, not a patch problem — and it’s a reminder that the security agent itself is now part of the attack surface.
Sources: BleepingComputer, SecurityWeek, Help Net Security
Vulnerabilities & Exposure
Microsoft Defender zero-day “RoguePlanet”. A time-of-check-to-time-of-use race condition in Defender’s real-time scanning engine (CVE-2026-50656) lets an attacker spawn a SYSTEM-level command prompt on fully patched Windows 10 and 11 machines. Microsoft has confirmed the CVE and says a patch is in development, but none is available yet, and a working proof-of-concept is already public — the researcher published it hours after June’s Patch Tuesday amid a disclosure dispute.
Oracle’s June Critical Patch Update. Oracle’s second monthly Critical Security Patch Update addresses 245 vulnerabilities, roughly 120 of them rated critical, with the heaviest concentration in Fusion Middleware (106) and E-Business Suite (55).
AI Impact on Security
The AI toolchain is becoming the new supply chain. Both the Mastra agent framework and the JetBrains AI-assistant plugins were poisoned to ride developer trust and harvest AI keys and secrets. As teams wire agents and AI coding assistants into production, the dependency graph and the underlying model/key plumbing are becoming first-class attack surface.
Telling agent activity from human activity is still unsolved. At RSAC 2026, CrowdStrike, Palo Alto and Cisco all shipped agentic-SOC tooling, but coverage analysis flagged a persistent gap: most SIEMs still can’t baseline or ingest agent-specific telemetry at volume, making it hard to distinguish a rogue or compromised agent from a normal one.
So what: “We have agents” is now table stakes. The unmet need is observing the agents — their identities, calls, and data access — which is fundamentally a visibility and data-foundation problem.
Cyberattacks & Breaches
FortiBleed: 30,000+ Fortinet firewalls compromised. An automated campaign is logging into Fortinet firewalls and VPNs using known and previously harvested passwords, then using each compromised device as a listening post to skim further credentials — no exploit required. Roughly 30,800 verified devices are affected across banks, hospitals, telecoms, energy providers and government agencies in more than 190 countries.
Mastra npm ecosystem backdoored across 140+ packages. Attackers took over the @mastra npm scope and, over roughly 88 minutes on June 16-17, republished more than 140 packages of the Mastra AI agent framework with a malicious “easy-day-js” typosquat dependency — a poisoned clone of the popular dayjs library. The postinstall dropper pulls a second-stage command-and-control implant that harvests browser data and crypto-wallet extensions; @mastra/core alone sees roughly 918,000 weekly downloads.
JetBrains Marketplace plugins steal AI API keys. Fifteen malicious IDE plugins spread across seven publisher accounts, with roughly 70,000 installs combined, quietly exfiltrated developers’ AI provider keys as plaintext over HTTP to a hardcoded command-and-control server. The campaign has been active since October 2025, with the newest plugin surfacing as recently as June 10; JetBrains has since purged the plugins and terminated the accounts.
So what: Two of today’s breaches involve poisoned AI tooling, and the Defender zero-day turns the security agent itself into the threat. The through-line: trusted software and security tooling are now the front line, and visibility into what your agents, build pipelines, and security tools are actually doing is the only durable answer.