Aug 14's Top Cyber News NOW! - Ep 1195

At a glance

Akira ransomware affiliate nearly pulled off a novel safe-mode persistence attack that failed due to virtual memory errors. Taiwan reported AI-assisted intrusions targeting government agencies and nuclear infrastructure. Microsoft's Patch Tuesday hit 419 vulnerabilities—a record spike driven by AI-accelerated vulnerability discovery—while Microsoft consolidates Copilot into a super app, raising data exposure risks.

Stories covered

How did Akira's safe-mode persistence technique nearly succeed?

What happened: An Akira ransomware affiliate rebooted a compromised server into Windows safe mode to disable EDR and Microsoft Defender in one move, then pre-staged a registry entry to keep the attacker's malware running through the reboot. The attack failed when the system ran out of virtual memory during the reboot, crashing the ransomware process before encryption could begin.

Why it matters: This demonstrates attackers are innovating against detection and response tools. However, the initial compromise vector—a VPN account with no MFA—was trivial. Without basic hygiene controls, sophisticated techniques don't matter.

What to do: Enforce MFA on all VPN accounts. Monitor for bursts of failed login attempts followed by successful access. Ensure EDR is deployed across every endpoint, not just a subset. Prioritize credential hygiene and account security over advanced detection tuning.

---

Is WhatsApp's new scam detection useful for consumers?

What happened: WhatsApp rolled out an optional, limited-beta scam alert feature using on-device machine learning to flag suspicious incoming messages from non-contacts. The model runs locally and doesn't report back to WhatsApp or Meta.

Why it matters: Pig butchering and credential theft via SMS remain endemic. Automated warnings help non-technical users identify obvious scams, raising the attacker's success threshold slightly without false positives deterring legitimate use.

What to do: Monitor for now. Educate users that scam detection is probabilistic—messages flagged as scams may still be legitimate, and messages that pass screening may still be malicious. Threat actors will likely adapt messaging patterns over time.

---

What does the White House's private sector hacking program mean for offensive security?

What happened: The U.S. President signed a national security memorandum directing private sector companies to support law enforcement and federal agencies in offensive hacking operations against transnational criminal organizations. Participating companies must sign contracts with DOJ or DHS and undergo vetting. Industry reactions are mixed, with some calling it progress and others warning of unchecked billable operations.

Why it matters: This effectively deputizes private sector cyber operators for offensive national security missions, creating a market incentive for offensive security talent and expanding the U.S. cyber warfare apparatus. It mirrors how private military contractors operate in kinetic domains.

What to do: Monitor for now. If your organization participates, ensure clear rules of engagement and governance structures exist. For practitioners in offensive roles, expect expanded contractor opportunities. Understand the strategic and diplomatic risks of private-sector offensive operations.

---

How did AI-assisted attackers compromise Taiwan's government?

What happened: Taiwan's Ministry of Digital Affairs reported an abnormal cyberattack starting July 20 that used open-source AI agents to automate a coordinated hacking campaign. The autonomous tool compromised at least 85 government user accounts, exfiltrated 2,500+ personnel records, then expanded to Taiwan's nuclear safety agency and seven energy companies. No confirmed attribution, but suspicion points to China.

Why it matters: This signals AI-assisted attacks are maturing and becoming faster than human-led campaigns. The speed and scale of the intrusion—crossing from government to critical infrastructure—demonstrates the compression of time-to-impact. Expect this pattern to repeat against high-value targets worldwide.

What to do: Audit your credential stores and cache for lateral movement risk. Monitor for rapid, bulk account compromises. Assume nation-state involvement until proven otherwise. Implement segmentation between government, regulatory, and critical infrastructure networks.

---

Why is Patch Tuesday breaking vulnerability management programs?

What happened: Microsoft released 419 security vulnerabilities in August's Patch Tuesday, including 62 critical fixes and 3 zero-days. This follows record-breaking months in June (206) and July (622), driven by AI's ability to discover and generate exploitable flaws at scale.

Why it matters: Vulnerability management teams cannot patch at AI speed. Threat actors use AI to reverse-engineer exploits within days. The CVSS/EPSS-only sorting approach fails because scores ignore your environment and critical assets. Vulnerability managers are operationally overloaded and burning out.

What to do: Stop prioritizing by CVSS alone. Map your critical assets, identify systems with access to them, then prioritize vulnerabilities on those systems. Implement automation for patching non-critical systems. Consider risk-based remediation frameworks that account for threat actor interest and your exposure window.

---

Should I trust Anthropic's Claude watermarking for EU compliance?

What happened: Anthropic announced plans to digitally watermark all Claude-generated content to comply with EU transparency rules. Watermarks persist even if Claude content is used for supporting tasks like spellcheck. Some worry this will attribute all employee work to AI when tools are heavily encouraged.

Why it matters: The watermarking mechanism (text-based, not image-based) is unclear—copy-paste behavior could strip watermarks, defeating the purpose. Anthropic's scraping of copyrighted material to train Claude, then watermarking outputs, highlights a transparency double standard.

What to do: Monitor for now. If you use Claude in corporate workflows, clarify with legal and compliance teams how watermarked content affects IP attribution. Don't assume watermarks are tamper-proof.

---

How are attackers exploiting Salesforce and ServiceNow portals?

What happened: A campaign called "city-forum" uses custom tools to scrape data from Salesforce Experience Cloud and ServiceNow customer portals configured with overly permissive guest access. The single-source server attack targets telecom, banking, fintech, software vendors, and public sector organizations. No vulnerability exploitation—only misconfiguration exploitation.

Why it matters: Poorly configured customer portals expose sensitive data to unauthenticated users. This is a recurring pattern across SaaS platforms. Blocking the attacker's IP address is ineffective; the threat actor can rotate infrastructure trivially.

What to do: Audit all public-facing Salesforce Experience Cloud and ServiceNow instances for excessive guest permissions. Disable unauthenticated API access where possible. Review sharing rules and portal configurations with your dedicated ServiceNow/Salesforce teams. Hunt logs for anomalous API calls originating from unknown IPs.

---

Why is Microsoft consolidating Copilot into a super app?

What happened: Microsoft announced a plan to merge consumer and business Copilot applications into a single "super app." Beginning September, Copilot Chat, AI coding, Copilot in Work, and new autopilot agents will unify into one interface with updated branding and web address.

Why it matters: Microsoft is baking AI deep into enterprise workflows to compete with ChatGPT, Gemini, and Claude. Copilot will gain access to legacy databases, ODBC connections, and third-party integrations. This expands the attack surface for data exfiltration if queries lack proper access controls.

What to do: Before connecting legacy databases or third-party systems to Copilot, conduct access control audits. Ensure Copilot inherits and enforces the original data source's authentication and authorization. Monitor for unusual query patterns. Don't assume M365-native security extends to external data sources connected via Copilot.

Key takeaways

• MFA failures and credential spraying remain the easiest path to breach; sophisticated post-compromise techniques are secondary to basic hygiene.
• AI is compressing the vulnerability-to-exploit timeline from weeks to days; CVSS-only remediation strategies are obsolete—prioritize by asset criticality instead.
• Misconfiguration exploitation (open portals, overly permissive sharing) is as dangerous as CVEs; audit SaaS portal access controls immediately.
• AI-assisted attacks are scaling and automating the full kill chain; assume nation-state capability when speed and coordination exceed typical human-led intrusions.
• Copilot and other AI super-apps will expose legacy data if connected without proper segmentation and access controls; treat them as untrusted intermediaries.

Topics covered

ransomware, Akira, safe mode persistence, MFA, WhatsApp, scam detection, offensive hacking, private sector, Taiwan, critical infrastructure, AI-assisted attacks, Patch Tuesday, vulnerabilities, Microsoft, CVSS, vulnerability management, Claude, watermarking, Salesforce, ServiceNow, Copilot, data exposure

Show notes generated from the live transcript using AI on Fri, 14 Aug 2026 18:11:54 GMT. Errors? Open the YouTube replay for the source of truth.