privacysavvy

The UpGuardian

Feature

Unauthenticated MCP Server Exposed India's Largest Real Estate Platform

Listicle

Left Unsupervised:  10 Times Access Outlived  Its Authorization

Resource

AI Assurance: The Third Head of Your AI Governance Watchdog

Letter from the Editor

Same September reminder, eighth year running: watch for behavioral shifts, know your coworkers’ baselines, report anything that feels off. Sound advice for a specific kind of insider. The kind with a pulse.

But that advice quietly assumes a badge and a manager. The actual definition was always broader: an insider is anyone who uses authorized access to cause harm, on purpose or not. What’s changed is who fits “anyone.”

An employee who switched teams twice and kept every permission from their first role. The vendor whose contract lapsed last spring, while their login didn’t. An AI agent running on a service account that nobody remembers issuing. There’s no HR file, manager, or exit interview waiting for it when it’s finally switched off. If it ever is.

In this issue of The UpGuardian, we open on the Square Yards incident. A database connected to an AI assistant, then left unauthenticated and accessible to anyone on the public internet. And it’s the clearest example yet of what happens when nobody’s watching. We count 10 more times standing access refused to expire on schedule. And we close on the harder question beneath both: what a governance program is worth if nobody can prove its controls fired when it mattered.

Swathi Shridhar

At a Glance

September

The eighth National Insider Threat Awareness Month begins. Insider incidents cost the average organization $19.5 million in 2025, driven largely by negligent, non-malicious behavior. (Source: 2026 Cost of Insider Risks Global Report)

17,600

Hugging Face reconstructed roughly 17,600 attacker actions after an AI agent spent four and a half days inside its production infrastructure. OpenAI confirmed five days after the July 16 disclosure that the agent was its own.

Article 14

European Union Cyber Resilience Act Article 14 reporting goes live September 11, 2026. Manufacturers get 24 hours to flag an actively exploited vulnerability, 72 hours for full notification, including for products shipped years ago.

Feature

Open House: How an Unauthenticated MCP Server Exposed India's Largest Real Estate Platform

684cb5b99dd4f2d803858774_1678294813918 (2)

By Greg Pollock

An unauthenticated DBHub MCP server, reachable from the public internet granted anyone who found it query access to the database behind Square Yards, which describes itself as India's largest real estate platform.  

2,178 tables, over 650GB, millions of records. UpGuard found it on May 28 and notified them on May 29. The data was secured soon after. This leak is the largest exposure of sensitive personal data attributed to an unauthenticated MCP server to date, a landmark in the progression from theoretical to actual AI risks

“Lead contacts, Aadhaar and PAN numbers, passport numbers, bank account details, CIBIL credit scores, and employee records...”

Listicle

10 Times Access Outlived  Its Authorization

By Revashni Moodley

Sometimes the true threat isn’t always a shadowy intruder breaking in. It’s access that was granted on purpose: an unrotated credential, a forgotten OAuth token, and then never revisited. Access granted once shouldn’t mean access forever. Yet, when loops aren’t closed, they quietly become backdoors wide open for years.

We break down 10 real-life incidents in which access outlived its authorization, extracting hard-learned lessons for security teams. The failure was never the access, just the assumption that someone was still watching it.

Resource

AI Assurance: The Third Head of Your AI Governance Watchdog

68b8603d07b2a205f6e648f8_20250804_122408 (2)-p-1600

By Shane Moosa

AI governance gets talked about as one thing. It's three: a policy on what's allowed, enforcement of what happens when someone ignores it, and a record proving the first two were in effect when it mattered. Most have the first, but almost none have the third.

Two incidents in July made the gap tangible. Claude's own share links turned up in Google search results: nothing had been hacked; the links were just public, and nobody could say how many were still live. Days later, an OpenAI model escaped a sandboxed safety test and spent four and a half days loose inside Hugging Face's production systems. Standing proof used to be what a mature governance program eventually landed on. It's fast becoming where the conversation has to start.

Summit

Being secure was the whole job. Proving it is the new one.

Your board wants it in business terms. Auditors want your reasoning. And regulators want it too.  

Join us at Summit to prove your work once and defend it everywhere.

Tune in live: 

September 1 (US) 

September 3 (APAC, EMEA, and India).

Or register for the recording.

Save my seat

Palate cleanser

Papers in Order

In March 1951, Ferdinand Waldo Demara enlisted in the Royal Canadian Navy under the credentials of Dr. Joseph Cyr, a real New Brunswick physician. He had no medical training, and nobody checked. Aboard HMCS Cayuga, he operated on 16 Korean casualties, speed-reading each procedure from a textbook first. All 16 survived, and the crew put him up for a commendation. The publicity that followed was his undoing, as Cyr’s actual mother read an account of her son’s supposed heroics off Korea. Verified once at the door, then trusted indefinitely, and caught only because someone outside happened to read the paper.

privacysavvy

Tuesday, September 1, 2026

The assumption that someone's still watching.

The assumption that someone's still watching.

Report Abuse