☀️ Good morning. Here's everything that happened in cybersecurity yesterday — in under 5 minutes.
Prompt injection took center stage as our new explainer landed on the leading vulnerability in every LLM-powered app, and the theme held all day: CrowdStrike's 2026 Threat Hunting Report declared AI "both the weapon and the target," clocking an 89% surge in machine-assisted attacks and patch windows shrinking to 48 hours. Microsoft pinned a wave of hotel Wi-Fi compromises on Russia's Midnight Blizzard, N-able scrambled to ship a real fix for an actively exploited N-central authentication bypass after its first patch was bypassed, and a legacy bug drained roughly $89 million from Coldcard Bitcoin wallets. Underneath it all, Iran-linked hackers pushed their assault on US water systems past Minnesota into at least six more states.
🔥 Top Stories
01 — Prompt Injection Is Now the No. 1 Vulnerability in LLM Applications
AI Security
Large language models now sit under customer-support agents, coding assistants, and internal search, and they share one structural weakness: they can't reliably tell developer instructions apart from user input or retrieved content — it's all just text in the context window. Prompt injection exploits exactly that, planting instructions inside what the model reads so it follows the attacker instead of the developer. OWASP ranks it the top risk for LLM applications, with documented cases spanning leaked system prompts, data exfiltration through connected tools, and unauthorized actions taken on a user's behalf.
The uncomfortable takeaway is that there's no single fix at the model layer today — indirect injection through poisoned documents, emails, or web pages makes the user the victim, not the attacker. Defenders should treat every input as untrusted, constrain what the model is allowed to do, isolate and sanitize its output, and put prompt injection scenarios into AI red-teaming as a first-class test case.
02 — Microsoft Ties Hotel Wi-Fi Hijacks to Russia's Midnight Blizzard
Nation-State
Microsoft says Russian state-sponsored actors — tracked as Midnight Blizzard — have been compromising hotel and hospitality Wi-Fi networks worldwide to harvest travelers' credentials and drop espionage malware on their devices. The campaign targets the public gateway layer, turning the network a guest trusts into the initial access vector, with stolen Microsoft account credentials as a primary prize.
For anyone traveling — especially government, defense, and executive staff — the action is old advice with new urgency: treat hotel Wi-Fi as hostile, route everything through a VPN, and lean on phishing-resistant MFA so a captured password alone gets the attacker nowhere. Security teams should watch for anomalous sign-ins originating from travel networks.
03 — N-able Ships Real Fix After N-central Bypass Exploited in the Wild
Vulnerabilities
N-able confirmed attackers exploited an authentication bypass in its N-central remote monitoring and management platform to gain remote administrative access — and reach the downstream customer systems those servers manage. The catch: the company's first patch was incomplete, and threat actors found a bypass. CVE-2026-18577 affects N-central builds prior to 2026.3.1.7, which shipped August 2 as the first unaffected version.
Because RMM platforms sit at the top of the management chain, a single compromised N-central server can cascade to every managed endpoint beneath it. Administrators should upgrade to build 2026.3.1.7 immediately, hunt for signs of unauthorized admin access, and rotate credentials on any server that lagged behind the fix.
04 — CrowdStrike: AI Is Now "Both the Weapon and the Target"
Threat Intelligence
CrowdStrike's 2026 Threat Hunting Report frames the year's defining shift bluntly: AI is simultaneously what attackers wield and what they aim at. The firm now assesses roughly 2.5 machine-generated signals for every human-triggered one, tracked an 89% surge in machine-assisted malicious activity, and warns that adversaries are weaponizing vulnerabilities faster than defenders can patch — collapsing the exploitation window to as little as 48 hours.
The report reframes AI security from a niche concern into an operational reality for every SOC. Defenders can't out-click AI-accelerated attackers, so the emphasis moves to automated detection, faster patch cadence, and protecting AI systems themselves as first-class assets rather than experimental tools.
05 — Legacy Bug Drains Roughly $89 Million From Coldcard Bitcoin Wallets
Cybercrime
A hacker drained nearly $89 million from Coldcard Bitcoin hardware wallets after exploiting a legacy bug in the devices, according to Infosecurity Magazine. The theft is one of the larger single-vector hardware-wallet losses on record and underscores that "cold storage" is only as safe as the firmware and legacy code paths beneath it.
Coldcard users should confirm their firmware is current, review guidance from the vendor, and — where feasible — migrate funds to unaffected configurations. More broadly, the incident is a reminder that self-custody shifts the security burden entirely onto the holder, legacy code and all.
📊 By The Numbers
$89 million — drained from Coldcard Bitcoin wallets via an exploited legacy bug.
89% — CrowdStrike's tracked surge in machine-assisted (AI-driven) malicious activity.
48 hours — how short the exploitation window has shrunk as AI accelerates attacks, per CrowdStrike.
6+ states — US water systems now targeted by Iran-linked hackers, expanding beyond Minnesota.
CVE-2026-18577 — the N-able N-central authentication bypass exploited after an incomplete first fix; patched in build 2026.3.1.7.
⚡ The Signal
Read yesterday's stories together and one line runs through nearly all of them: AI has stopped being a side conversation in security and become the terrain the fight happens on. Our prompt-injection explainer describes the structural flaw at the model layer; CrowdStrike's report quantifies what that flaw looks like at scale — machines outnumbering humans in the signal stream, and a 48-hour window between disclosure and exploitation. When the same technology is both the tool attackers reach for and the asset they aim at, "AI security" stops being a specialty and becomes a baseline.
But the day's other headlines are a useful corrective against AI tunnel vision. Midnight Blizzard is still stealing credentials off hotel Wi-Fi. N-able's crisis is a patch-bypass on a management platform. Coldcard's $89 million loss traces to a legacy bug, not a clever prompt. The fundamentals — patch fast, treat untrusted networks as hostile, use phishing-resistant MFA, mind your legacy code — didn't get less important because AI got more capable. They got harder to execute inside a 48-hour window.
The practical synthesis for defenders: the attack surface is widening at both ends. You now have to secure the AI systems you deploy and keep pace with AI-accelerated attacks against everything else. The organizations that treat their LLM applications as security-critical — and their patch cadence as a race, not a checklist — are the ones the next report won't be written about.
🔍 What You May Have Missed
OpenAI Reveals How Criminals Used ChatGPT to Run Scams — OpenAI banned a coordinated network, likely tied to Cambodian scam compounds, that used its models to build fake personas, translate messages to targets, and run daily fraud operations.
Brinks Home Discloses Data Breach as Hackers Leak Files — The physical-security firm says attackers leaked files but that its alarm monitoring and system functionality were not affected.
Korea's Largest Telco KT Fined About $38 Million After Femtocell Campaign — Regulators penalized KT over a year-long breach linked to femtocell compromise, one of the steepest telco security fines to date.
📅 What to Watch
US Water Cyberattacks Extend Beyond Minnesota to at Least 6 Other States — Iran-linked hackers are reportedly hitting water systems in Michigan, South Dakota, Georgia and more; expect the list of affected states to keep growing.
KindaRails2Shell Threatens Ruby on Rails Apps (CVE-2026-66066) — A critical Rails flaw lets an attacker slip a booby-trapped file past image uploads to read server secrets and, in some cases, take full control; patch as details spread.
Hugging Face Diffusers Flaws Could Let Model Repositories Execute Arbitrary Code — Three high-severity bugs bypass the trust_remote_code safeguard, putting the AI model supply chain at risk for anyone loading community repositories.
Pass the Passkey: A Novel Attack Surface in Passwordless Authentication — Unit 42 shows how implementation gaps — relying parties skipping the User Verified flag — can quietly reduce passkey MFA to a single factor.
Stay sharp. Stay ahead.
Till next time,
The CyberSignal Team