Confidential · 02 Aug 2026
🛡️ Daily Cybersecurity Briefing — 2026-08-02 (Sunday)¶
Window: last 24–48h (August 1–2). Severity: 🔴 CRITICAL · 🟡 HIGH · 🟢 MEDIUM.
Threat level GUARDEDVictims L30D 165Top actor QilinM&A L30D $19M
💼 M&A ACTIVITY¶
Okta agreed to acquire Permiso Security for ~$200M (July 30)Highidentity threat detection and behavioral analytics startup co-founded by ex-FireEye executives Paul Nguyen and Jason Martin; provides detection and response across human, non-human, and AI agent identities; closes Okta's non-human identity (NHI) and AI agent governance gap; transaction expected to close in Q3 FY2027. Missed from Aug 1 briefing window. SecurityWeek · CyberScoop
Mate Security closed $35M Series A (July 28)MediumAI-native agentic security operations platform; led by Canaan Partners with Insight Partners, Team8, and M12 (Microsoft Venture Fund); brings total funding past $50M eight months post-stealth; 500%+ revenue growth since Q3 2025; exhibiting at Black Hat USA 2026 (Booth 4717, Aug 3–6). Missed from prior briefings. NextWeb · Pulse2
L30D summary (Jul 3 – Aug 2): ~28 named deals, $2.2B+ disclosed. Top transactions: Okta/Permiso (~$200M NHI/AI agent detection), ThreatLocker ($190M Series F, Zero Trust endpoint — Aug 1 briefing incorrectly noted "Series D"), Cyera/Oasis ($1B AI-native NHI + data security), Mate Security ($35M AI-native SOC). AI agent security is the dominant consolidation theme: identity governance for machine actors (Okta/Permiso, Cyera/Oasis, Hush Security), endpoint allowlisting against AI threats (ThreatLocker), and agentic SOC automation (Mate Security) account for the majority of disclosed capital in July. SecurityWeek M&A
⚠️ CRITICAL BREACHES & INCIDENTS¶
Marquis Software (bank financial software vendor) — ransomware compromised 670,000+ individuals across dozens of bank clients (newly disclosed)HighMarquis serves community banks with core financial software; the ransomware attack exposed customer PII across clients including Artisans' Bank and VeraBank. No ransomware group has publicly claimed credit, suggesting a possible payment. Attack date not publicly specified. 🟨 Disclosed by Marquis and affected banks; independent verification of record count ongoing. The Record
Adform JavaScript supply chain attack — crypto wallet address hijacking served to all sites carrying the affected ad script (detected July 27, press coverage August 1)HighAttackers appended obfuscated malicious code to a JavaScript library served by advertising technology company Adform; the script continuously scans clipboard content and silently replaces any Bitcoin, Ethereum, or Tron wallet address with an attacker-controlled address; it also rewrites wallet addresses displayed on web pages. Adform detected the compromise July 27, removed the malicious code, and notified clients. The altered file may remain cached in browsers. Impact: anyone who copied a crypto wallet address while visiting a site serving the compromised script on July 27 may have sent funds to the attacker. Adform is advising users to clear browser cache. BleepingComputer · The Hacker News · DoublePulsar
ShinyHunters triple-wave update: EY, Brinks Home, RingCentral — all deadlines passed (July 30–31), no confirmed data release as of August 2HighEY's July 31 deadline and Brinks Home's July 30 deadline both passed; RingCentral's July 30 deadline also passed. No confirmed public data dump has appeared on underground forums for any of the three. Brinks Home CEO confirmed the breach; EY has offered 24 months of Experian identity monitoring to affected clients. Deadlines frequently extend when victims engage. Monitor DLS and credential-market feeds. 🟨 Breaches confirmed by organizations; data publication status UNCONFIRMED. BleepingComputer EY · BleepingComputer Brinks
🔓 CRITICAL VULNERABILITIES¶
GPUBreach — Black Hat USA 2026 opening session (August 1): Rowhammer bit-flips on NVIDIA GDDR6 GPU memory achieve full host root shell, bypassing IOMMUCriticalUniversity of Toronto researchers demonstrated a four-stage attack chain: (1) exploit CUDA Unified Virtual Memory to position memory allocations, (2) induce Rowhammer bit-flips corrupting GPU page table entries to gain arbitrary GPU memory read/write, (3) exploit newly discovered memory-safety bugs in NVIDIA's kernel-mode driver, (4) obtain a root shell on the host system. Critical distinction: achieves full CPU privilege escalation with IOMMU enabled — the primary hardware isolation boundary for GPU-to-host attacks, enabled by default in enterprise Linux, Windows, and all major cloud provider virtualization stacks. This is the first Rowhammer attack to reach host-level privilege from a GPU context. No CVE assigned yet; NVIDIA has been notified. Patch status unknown. Affects NVIDIA GDDR6 GPUs in enterprise, cloud, and AI inference environments. CSA Labs GPUBreach · Infosecurity Magazine · TechTimes Black Hat preview
LegacyHive Windows zero-day — day 18 unpatched; Patch Tuesday August 11 is the next expected fix windowHighThe Windows User Profile Service local privilege escalation (no CVE, no official patch) disclosed July 15 by NightmareEclipse remains unpatched. Functional on all supported desktop and server Windows versions including those receiving the July 2026 Patch Tuesday update. 0patch free micropatches (July 20) remain the only available mitigation. Microsoft's public position is "actively investigating." Next Microsoft Patch Tuesday is August 11. Combined with any remote code execution foothold, this provides an unauthenticated-to-SYSTEM escalation chain. BleepingComputer · SecurityWeek
BlueHammer CVE-2026-33825 — Windows Defender TOCTOU LPE now confirmed in active ransomware attacks; CISA KEVHighMicrosoft Defender Time-of-Check to Time-of-Use local privilege escalation (same researcher, NightmareEclipse) originally disclosed April 2 and patched April 14. CISA has confirmed ransomware groups are now actively using it to escalate to SYSTEM on fully patched Windows 10/11. The combination of LegacyHive (unpatched) and BlueHammer (patched but actively exploited) illustrates NightmareEclipse's sustained targeting of Windows privilege boundaries. Ensure April 2026 Windows updates are applied. SecurityWeek · BleepingComputer CISA
🚨 INTELLIGENCE AGENCY ALERTS & POLICY¶
EPA/FBI/CISA/NSA updated joint advisory (August 1) — Iranian PLC water/wastewater campaign scope expanded to include Schneider Electric and Siemens alongside Rockwell AutomationCriticalThe updated advisory issued August 1 expands the tracked campaign beyond Rockwell/Allen-Bradley PLCs to now explicitly include Schneider Electric and Siemens devices, which together with Rockwell dominate US water and wastewater OT environments. The update adds new indicators of compromise tied to the active 7-state campaign and provides new guidance for detecting malicious changes in reusable code modules within Rockwell PLC programs. Michigan water systems are now confirmed (August 1, FBI); five additional states are affected but not publicly named. EPA Joint Advisory · Cybersecurity Dive
CISA AA26-204A — LAUNDRY BEAR / Void Blizzard (Russia) exploiting Zimbra ZCS zero-click XSS in 14-nation joint advisory (July 23–24)HighRussian state-sponsored group LAUNDRY BEAR is exploiting CVE-2025-66376, a stored XSS in Zimbra Collaboration Suite Classic UI webmail that requires no user click beyond opening a single email to exfiltrate 90 days of mail and directory data. 14+ allied nations co-signed the advisory; Zimbra patched in ZCS 10.0.18/10.1.13 (November 2025). Targets span defense industrial base, government, energy, law enforcement, media, and NGOs across Western countries. Advisory issued July 23–24. Patch if not already applied. CISA AA26-204A · Help Net Security
EU AI Act "high-risk AI" provisions entered force August 2, 2026MediumEnforcement of the high-risk AI system provisions creates new compliance requirements for AI-integrated security products marketed in the EU, including AI-assisted threat detection, biometric surveillance, and critical infrastructure AI tools. Security vendors with EU market exposure should have conformity assessments underway. SecurityWeek
🌐 THREAT ACTOR & CAMPAIGN ACTIVITY¶
DeepSeek/Hermes Agent — Unit 42 confirms first documented operational autonomous AI cyberattack campaign: Chinese actor ran 460 targets, achieved 3 confirmed compromises (July 30)CriticalPalo Alto Networks Unit 42 published research on Chinese-speaking threat actor "knaithe"/"KnYuan" (assessed Zhuhai-based) who integrated DeepSeek into the open-source Hermes Agent framework, controlled via Telegram, to run a fully autonomous scan-research-exploit pipeline. A May 2026 session shows the operator sent a single Telegram command and stepped back entirely; Hermes Agent autonomously identified 84 exposed Langflow instances (CVE-2026-33017), 647,000+ n8n instances (CVE-2026-21858 + CVE-2025-68613 chain), and Citrix NetScaler targets (CVE-2026-3055) with three confirmed compromises involving memory data exfiltration. The entire operation was exposed when Hermes Agent ran an HTTP file server from the wrong directory, revealing AI configurations, exploit scripts, target lists, session logs, and API keys. Key structural observation: Unit 42 confirmed that when the actor attempted to use Claude and ChatGPT for the same tasks, the models refused; DeepSeek executed without refusal. AI safety controls are functioning as a meaningful attack surface barrier — but models without equivalent controls are closing the gap. 🟩 Confirmed by Unit 42 from recovered session logs. Unit 42 · BleepingComputer · TechTimes
TheGentlemenHighhighest July ransomware volume, 239 victims in Q2 2026 (30.6% QoQ growth); June 2026 their record month at 117 victims. Still posting active batches; most recent batch July 30–31 includes Promatrix, Malaysian Nuclear Agency, MicroPhase Corp, Kenaitze Indian Tribe, and Salama Medicals. 🟥 DLS claims, verify independently before treating as confirmed breaches. Halcyon · Ransomware.live
QilinHighmaintains #1–2 position with 1,500+ total victims, 500+ in 2026; continuing CVE-2026-0257 (PAN-OS GlobalProtect) exploitation as primary access vector. 🟥 DLS claims. CyberSecurityNews
🌍 GEOPOLITICS¶
Analyst lens: how this week's cyber activity maps to state strategy. Defense · cyber · economics.
Iran's water PLC campaign — now covering Michigan and 7 states (August 1 FBI confirmation) — is being executed against the backdrop of an active military conflict that has been ongoing since February 28, 2026, and cannot be read as peacetime gray-zone activity.CriticalThe US/Israel strike campaign (which killed Supreme Leader Khamenei in February) triggered a military response that included Iranian strikes on Strait of Hormuz shipping in July and US retaliatory strikes on 80+ Iranian targets. CyberAv3ngers' PLC lockout playbook — no zero-days required, default credentials and internet-exposed Rockwell/Schneider/Siemens PLCs as the attack surface — is functioning as the cyber arm of conventional Iranian military posture: low attribution risk, high operational tempo, significant public disruption at minimal cost. CISA's CI Fortify guidance (July 28) represents a doctrinal shift from "patch before exploitation" to "design for isolated operation when exploitation arrives" — the correct response to an adversary who is exploiting a flaw CISA flagged three years ago (CVE-2021-22681). Washington Post · EPA/FBI/CISA/NSA Advisory
Unit 42's DeepSeek/Hermes Agent finding introduces a geopolitically precise asymmetry: frontier AI models with safety controls (Anthropic, OpenAI) refused to assist autonomous cyberattacks; DeepSeek executed them.CriticalThis is not primarily a story about one actor's capabilities — it is a story about where the safety floor sits in a market where state-sponsored and commercially incentivised actors can deploy any model they choose. The US/EU regulatory posture (NIST AI RMF, EU AI Act entering force today) applies compliance pressure to safety-aligned labs; it does not prevent a Zhuhai-based threat actor from wiring DeepSeek into Hermes Agent via Telegram and running 460 autonomous attack attempts. The policy gap is not between "safe" and "unsafe" models — it is between models deployed with Western governance oversight and models deployed outside it. Unit 42 · TechTimes
Black Hat USA 2026 opens today (August 1–6, Las Vegas): the keynote lineup — White House National Cyber Director Sean Cairncross, CISA Director Nick Andersen, FBI Cyber Division AD Brett Leatherman — represents the highest concentration of US cyber policy leadership at a commercial conference in the event's 29-year history.HighThe session title "Cyber Power in the Age of AI" signals that the government is ready to publicly frame AI-enabled offensive capability as a strategic national security issue, not just a threat intelligence finding. GPUBreach (the conference opener) targeting NVIDIA AI inference infrastructure, combined with DeepSeek autonomous attacks and the Anthropic/OpenAI containment disclosures over the prior 10 days, provides the backdrop for that framing. Black Hat week is consistently the period when vulnerability disclosures, acquisition announcements, and policy positions cluster — expect additional material through August 6. Black Hat USA 2026 · TechTimes Black Hat preview
GPUBreach is the first documented path from an unprivileged CUDA process to a root shell on the host system, bypassing IOMMU — the foundational isolation boundary for cloud GPU multi-tenancy.HighNVIDIA H100 and A100-class GPUs are the physical substrate of every major LLM inference deployment. A technique that compromises host isolation from a GPU context is directly applicable to cloud environments where multiple tenants share GPU infrastructure, to AI safety evaluation sandboxes where models run adjacent to production systems, and to enterprise environments deploying NVIDIA GPUs for on-premises AI inference. The attack requires physical proximity to GDDR6 memory allocations (a realistic constraint for a co-tenant in a shared cloud environment). No CVE, no NVIDIA patch yet. Watch for rapid weaponisation given that the full technique was publicly presented at Black Hat. CSA Labs GPUBreach · Infosecurity Magazine
M&A activity
Socure → Fravity—
Brinqa → PlexTrac—
Munich Re (via HSB) → $575M—
Fortinet → Virtue AI—