Ten Hours From First Touch to Encryption: The SOC Clock Just Broke

Week of 7 September 2026 | Compiled by Team Paringa

Three stories this week point the same direction: attacker tempo is now set by machines, and the detection and response models most organisations bought were priced for human tempo.

What Happened This Week

🔴 AI agents ran an entire ransomware intrusion in under 10 hours

Palo Alto Networks Unit 42 published an incident report on 2 September describing an intrusion that took less than 10 hours end to end. The same attack pattern executed by human operators typically runs about two weeks. The human attacker admitted to negotiators that they used frontier models and agentic attack frameworks, with AI agents handling each step of the chain.

The sequence matters more than the headline. Recon agents mapped the target before entry. Initial access came through an exposed public API endpoint. Once inside, an automated recon agent mapped internal microservices while subagents scraped code repositories for hard-coded tokens and service passwords. Those tokens opened the secret-management system, which yielded master administrative credentials and root access.

Specialist pivot agents then validated access across cloud, identity, CI/CD, container and SaaS planes. The attacker hijacked CI/CD workflows to steal cloud access keys and turned the victim's own cloud AI services into post-compromise infrastructure, burning the victim's compute while hiding orchestration traffic inside legitimate activity. On the way out, an agent generated an 80-page report on the organisation's security failings, documenting dozens of exploited findings.

Unit 42's own framing is the part worth quoting to a board: there was no zero-day and no elite tradecraft. The differentiator was AI-assisted operational efficiency. Their recommendation is automated playbooks that simultaneously revoke credentials, terminate OAuth sessions, freeze CI/CD pipelines and isolate cloud accounts across every operational plane, plus treating AI as core infrastructure with an inventory of every model endpoint, API key, MCP gateway and tool integration.

Sources: The Register, Unit 42

A 10-hour intrusion invalidates most MDR SLAs written in the last three years. If your contract promises triage in 15 minutes and analyst escalation in an hour, you are still inside the attack window when the encryption starts. Two questions for any provider this quarter: what is your mean time to contain, not detect, and which containment actions execute without a human in the loop. Second point for the CFO in the room: the attacker spent the victim's own AI compute budget as attack infrastructure. That shows up as a cloud bill anomaly before it shows up as a security alert, which is exactly why cost telemetry and security telemetry belong in the same review.

🟡 Working zero-days dropped for CrowdStrike, Kaspersky and Avast in one week

An anonymous researcher using the handle "Nightmare Eclipse" published a CrowdStrike Falcon privilege-escalation exploit called FalconFlank on 4 September. It spawns a command prompt with SYSTEM privileges by abusing Falcon's Office malicious macro remediation feature, and the researcher says it works on fully updated Windows 11 25H2 and Windows Server 2025. No CVE has been assigned.

The same week brought privilege-escalation zero-days for Kaspersky Antivirus for Endpoint (HardBreacher) and GenDigital Avast Antivirus (PrettyPrague), plus a denial-of-service exploit for Nvidia (GreenSection). Kevin Beaumont confirmed the privilege-escalation exploits are real and functional. CrowdStrike told BleepingComputer it is investigating and advised customers to disable the Microsoft Office File Suspicious Macro Removal Windows policy setting, pointing to a tech alert that sits behind support-portal login rather than in public.

This researcher has been running a sustained campaign since April, with prior drops against Microsoft Defender, BitLocker and multiple Windows components. Several of those remain unpatched. Microsoft's response has been to warn of legal action, which has not slowed the release cadence.

Sources: BleepingComputer

The security agent is now a privilege-escalation primitive. Anything running as SYSTEM with a remediation function that touches user files is a lateral movement opportunity, and that applies to every EDR on the market, not just the three named this week. Practical action: confirm whether the Office macro remediation policy is enabled across your estate, and check whether your SOC has detection content for abuse of your own agent's remediation paths. Also worth noting for anyone running a single-vendor security stack, when the agent itself is the exploit, vendor telemetry is the last place the alert will come from.

🟡 5,400 hacked sites serve ClickFix payloads from blockchain smart contracts

Netskope researchers tracked more than 5,400 compromised websites, mostly WordPress and PrestaShop, injected with a script that pulls its next-stage payload from a smart contract on the BNB Smart Chain testnet. The technique, EtherHiding, puts the payload somewhere no takedown notice reaches. The operator can swap the payload at will and the testnet endpoint is free to use.

The visitor sees a fake CAPTCHA and is told to open the Windows Run dialog and paste a PowerShell command. Later in the campaign the operator swapped the ClickFix payload for a WebRTC data-channel stager that creates a peer connection, hand-writes its own session answer so no real handshake ever occurs, then opens an encrypted channel to the attacker. Received JavaScript is assembled in browser memory and executed without touching disk.

The campaign uses over 300 infected sites per day. Roughly 400 sites called the testnet RPC endpoints daily in August, peaking at 536. Netskope published the full RPC endpoint pool for blocking and recommends monitoring for non-web UDP traffic associated with WebRTC.

Sources: BleepingComputer, Netskope

Two detection gaps here. Blockchain RPC calls from a browser look like nothing in most network baselines, and a WebRTC data channel with no signalling server does not match any legitimate application pattern your team has modelled. Add the BSC testnet RPC pool to the blocklist this week, it is a five-minute change. The wider point for anyone running user awareness training: ClickFix works because it asks the user to do the attack themselves, which means every control that assumes a malicious download is watching the wrong step.

What CISOs Are Dealing With

The gap this week is between attacker tempo and operational tempo. Unit 42's 10-hour intrusion was not technically sophisticated, it was fast, and speed came from delegating execution to agents that monitor, evaluate, act and re-plan without waiting for an operator. Meanwhile the tooling that was supposed to buy time is itself becoming an attack surface, and delivery infrastructure is migrating to places that cannot be seized.

The practical consequence is that detection budget without containment automation buys very little. Most enterprise SOCs can see the events from the Unit 42 case in their logs. What they cannot do is revoke credentials, kill OAuth sessions, freeze pipelines and isolate cloud accounts in the same minute across identity, cloud, CI/CD and SaaS planes, because those are four consoles owned by three teams. The organisations that close that gap first will not be the ones with the most telemetry, they will be the ones who rehearsed the cross-plane containment play.

Three Things Worth Watching Next Week

1. CrowdStrike CVE assignment and patch for FalconFlank - The tech alert is currently customer-only and there is no CVE. Watch for public disclosure and whether the macro remediation feature is redesigned rather than just toggled off.

2. Copycat agentic intrusions - Unit 42 has published the pattern. Expect other IR firms to report similar sub-24-hour cases within weeks, and expect cyber insurers to start asking about containment automation at renewal.

3. ShinyHunters and the McKesson extortion deadline - The group claims 284 million customer records and a demand reported around $55 million. How that resolves sets the tone for large-scale data extortion pricing through Q4.

Questions? hello@paringa.ai

Paringa builds managed security with vCyberiz around one question: how fast can you contain, not how much can you see. If your MDR contract still measures success in detection time, we should talk. | 7 September 2026 | © Paringa Ltd