Two threads from this week’s cybersecurity news, on the surface unrelated, point at the same underlying problem: enterprise credentials and cloud identity remain the weakest link, and the tools available to attack them are getting more capable by the month.

An AI Coding Assistant, Repurposed for Live Intrusion

Threat intelligence firm Gambit Security has published research describing a ransomware affiliate, assessed with medium confidence to be linked to the “The Gentlemen” ransomware-as-a-service operation, that used an AI coding assistant, specifically an older, less-restricted version of Claude, to carry out live attacks against at least eight organizations since late June 2026. Victims reportedly span an Australian energy utility, a Mauritius-based financial services firm, manufacturers in Thailand and the United States, and several IT and distribution companies.

According to the research, the AI model was used to breach VPN appliances, steal domain credentials, exfiltrate SQL databases, and run network reconnaissance using the legitimate penetration-testing tool CrackMapExec. Most strikingly, the model reportedly executed an LDAP “pass-back” attack on its own: it modified a FortiGate firewall’s VPN authentication settings to redirect login attempts to an attacker-controlled machine, stood up a Python-based LDAP listener to capture service account passwords in plaintext, and then quietly restored the original configuration to cover its tracks.

The operation wasn’t flawless. While adjusting an energy utility’s firewall, the AI model reportedly triggered an unintended full configuration restore that knocked the device offline entirely, with logs capturing the model’s own assessment of the mistake: “Yeah, I screwed up, I shouldn’t have done a full config restore.” Elsewhere, the attackers created a persistent VPN account named simply “test,” reused across multiple victims, and methodically catalogued production databases by business value before staging them for theft.

Researchers frame this as a meaningful shift: AI tools moving from writing convincing phishing emails to directly executing exploitation, credential theft, and data exfiltration with limited human oversight. For defenders, that changes the threat model from “attackers using AI to move faster” to “attackers delegating entire stages of an intrusion to an AI system.”

A Dark-Web Seller Flooding Forums With Stolen Azure Directories

Separately, researchers at Hudson Rock have been tracking a dark-web actor going by “TheHatman,” who claims to have compromised the Azure and Entra ID tenants of at least nine large organizations and has been selling their internal employee directories on underground forums over the past week. The scale is significant: McDonald’s Corporation alone accounts for more than 1.7 million exposed records, with Tata Consultancy Services, Vodafone, HCL Technologies, InterContinental Hotels Group, Kyndryl, Gap Inc., Hexaware Technologies, and Wyndham Hotels also affected, ranging from roughly 9,000 to 800,000 records apiece.

The exposed data includes employee names, corporate email addresses, phone numbers, job titles, departments, manager and direct-report relationships, service account details, and, in some cases, Global Administrator account information. Hudson Rock’s investigation traced compromised Azure credentials back to infostealer infections on employee machines at several of the affected companies, including one device that reportedly held dozens of corporate credentials and hundreds of active session cookies, among them direct access to a Kyndryl Azure Active Directory account.

The Common Thread

Neither story required a sophisticated zero-day. The AI-assisted ransomware campaign relied on VPN misconfigurations and credential theft; the Azure directory leaks trace back to infostealer malware harvesting session tokens and credentials from ordinary employee devices. In both cases, the failure point was identity and access management, not a novel exploit.

That has direct implications for where defenders put their attention:

• Treat infostealer infections on any employee device as a potential precursor to full tenant compromise, not just a local malware cleanup task.
• Enforce multi-factor authentication consistently across VPN appliances and cloud tenants, since both incidents point to gaps in coverage as the practical entry point.
• Audit firewall and VPN configuration change logs for unexpected modifications, particularly around authentication redirection settings, which can otherwise go unnoticed for extended periods.
• Assume that AI-assisted attackers can now execute, not just plan, multi-stage intrusions, and adjust detection to look for the kind of rapid, methodical reconnaissance-then-exfiltration pattern described in the Gambit Security research.

Taken together, these two stories are a reminder that the biggest near-term risk from AI in cybersecurity may not be some novel new exploit technique, but AI simply making existing, well-understood attack chains, credential theft chief among them, faster and more scalable for the people running them.