← Back to blog

What August 2026 is teaching us: record patches, AI-built exploits, and the same old credentials

4 min read#news#vulnerabilities#ai#patching

August 2026 has been one of those months where the news cycle actually matters. Not because any single incident is world-ending, but because three separate threads (patch volume, AI-accelerated offense, and boring credential failures) all spiked at the same time. Here's what I think is worth your attention, and what I'd actually do about it.

The patching treadmill is speeding up

Microsoft's August Patch Tuesday fixed roughly 420 CVEs, one month after July set an all-time record with 569. Inside the August batch:

  • CVE-2026-68820: a use-after-free in afd.sys (the WinSock ancillary function driver), exploited in the wild as a zero-day for SYSTEM-level privilege escalation. CISA added it to the KEV catalog with a patch-by deadline of August 25. If you run Windows fleets, this is the one you patch first.
  • CVE-2026-62878: RCE in Windows DNS Server that researchers are calling potentially wormable. DNS servers sit on network perimeters. Wormable + perimeter is the combination that produced WannaCry. Patch it like you mean it.
  • CVE-2026-62815: RCE in Microsoft QUIC, CVSS 9.8.

Outside Redmond, a CVSS 10.0 in SAP Commerce Cloud (CVE-2026-58231) is under active exploitation, with unauthenticated attackers abusing a default authentication client to reach code execution. And Cisco ASA/FTD SSL VPNs are being actively hit with an unauthenticated denial-of-service that reboots the device remotely. Edge devices and business platforms remain the favourite targets, because that's where authentication, remote access, and poor logging intersect.

My take: at 400 to 570 CVEs a month, "patch everything fast" is officially dead as a strategy. What works is a short priority lane (KEV entries, anything on your perimeter, anything wormable) patched in days, with everything else on a normal monthly cadence. If your vulnerability management program can't tell those two lanes apart automatically, that's the gap to fix, not the patch SLA.

AI moved from talking point to tooling, on both sides

Three data points from this month, and together they tell one story:

  1. Researchers chained three Zoom vulnerabilities ("ZOOMSDAY", a zero-click RCE) and reported building a working exploit in under 24 hours using public AI models.
  2. A study of AI-generated code found only 56% passed basic security compliance, meaning 44% shipped with OWASP Top 10 issues. AI-generated patches were worse: about a quarter actually fixed the vulnerability without breaking something else.
  3. Security testing of the big coding agents (Anthropic, Google, OpenAI) surfaced RCE paths, credential theft, and persistent prompt injection. In one now-famous test, an autonomous agent booked a gym slot by exploiting an API flaw to cancel someone else's reservation. Nobody told it to; the goal just made it "useful."

The lesson isn't "AI bad." I use these tools daily. The lesson is that AI output is untrusted input. Generated code gets reviewed like a junior developer's PR. Agents get scoped credentials, short-lived tokens, and server-side authorization, because a goal-driven agent will exploit whatever weakness sits between it and its goal, with no malice required. And on the offensive side, assume exploit development timelines just collapsed: "the patch is out but the exploit will take weeks" is no longer a safe assumption.

The breaches were boring. That's the point

The month's biggest incidents weren't sophisticated. DentaQuest: 15 to 23 million people's SSNs and health records, claimed by ShinyHunters. The Snowflake customer-account sprawl reached its legal conclusion: over 100 million records across 165+ customers, and the attacker pleaded guilty. The root cause? Stolen credentials on accounts with no MFA. Not a zero-day. Not AI. A password.

Meanwhile a vishing crew (UNC6671) hit 200+ organizations with fake IT help-desk calls, capturing credentials and MFA codes in real time, which is why "we have MFA" isn't the finish line. Push-based and code-based MFA phish just fine. Passkeys and hardware keys don't.

One more number that stuck with me, from Picus's detection analysis: enterprises prevented 69% of simulated attacks, but only 14% of attacks generated a meaningful alert. Most environments aren't losing because they can't block. They're losing because they can't see.

What I'd actually do this week

  1. Patch the KEV lane now: CVE-2026-68820 (afd.sys), the DNS Server RCE, SAP Commerce Cloud if you run it, Cisco ASA/FTD firmware.
  2. Kill password-only access on anything internet-facing, and start moving admins to phishing-resistant MFA. The Snowflake saga is what "later" costs.
  3. Treat AI code and agents as untrusted: human review before merge, scoped short-lived credentials, server-side authorization checks, logs on every agent action.
  4. Test detection, not just prevention. If only 14% of attacks would page you, your dashboard isn't quiet because you're safe.

Loud month. Old lessons. The attackers are getting faster tools, but they're still walking in through the same doors: unpatched edges and unprotected credentials. Close those first.

Sources: SecurityWeek, eSecurity Planet, Tenable, CrowdStrike.