Shadow AI Epidemic: Uncovering Agents on the Endpoint, British Library Breach, & News – Amit Assaraf – ESW #475

Adrian Sanabria

1. FUNDING/M&A courtesy of the Security, Funded newsletter, issue 259 – In Fraud We Trust

   VIBE CHECK

   Is the "give every defender an agent" thesis the real security unlock, or is it the wrong bottleneck?

   • 46% - No - most orgs can't operationalize it fast enough to matter
   • 23% - It's vendor-driven urgency, not a real gap
   • 23% - No - triage judgment is the bottleneck
   • 8% - Yes - coverage was always the limit

   FUNDING

   • Socure, a United States-based digital identity verification and fraud prevention platform, raised a $156.0M Series E from Summit Partners.

   ACQUISITIONS

   • NetSPI and Synack reportedly merge
   • HOT RUMOR: Proofpoint in talks to acquire Varonis
   • ClickHouse, a database/data warehousing company, acquired RunReveal, a security log management vendor for an undisclosed amount.
   • Minimus (formerly Gutsy) and there assets thereof, an Israel-based secure-by-design container image company, was acquired by Echo for an undisclosed amount. Minimus had previously raised $51.0M in funding.

   SHUTDOWNS

   • Minimus, an Israel-based secure-by-design container image company, shut down operations. Its assets were sold to Echo (see above). Minimus had previously raised $51M in funding.
2. HOT TAKES: Rumors of the Vulnpocalypse Have Been Greatly Exaggerated

   Jeremiah and Robert have been doing some very interesting research over at Root Evidence. This webinar and the accompanying report share some interesting findings.

   The most significant hot take is that the popular ZeroDayClock.com's analysis was incorrect. The key miscalculation was that time-to-exploit was using the first sign of exploitation as point A and when MITRE added the CVE to NVD as point B. The calculation does a better job of measuring how fast MITRE moves than measuring attacker exploit development speed.

   ZeroDayClock.com has been wiped clean and appears to be under construction. It's unclear if this is directly tied to the release of this report, or if it is coincidental. Based on my conversations with folks in the vuln research community, I'm fairly certain Sergej Epp was aware of the alleged shortcomings of his calculations.

   The silver lining is that this is what science looks like! Someone publishes something and someone else attempts to reproduce their findings and finds issues. All published out in the open, so that anyone can help and contribute to a better understanding of the state of vuln mgmt and exploitation.

3. VULNERABILITIES: Everything I own, owned

   On a whim, the author points AI at all the accessories connected to his computer to see if it can find significant vulnerabilities. Nothing survived unscathed.

   While none of the vulnerabilities are groundbreaking, it's a sobering experiment towards validating a significant and common AI hype claim: that you can point an LLM at damn near anything and come away with vulnerabilities. There was no cherry picking here - these were literally just thing things on the author's desk.

   I think this says a lot about the potential of future targeted attacks. Currently, the evidence we have says that opportunistic attacks are the majority, but if someone really wants to compromise a specific organization, it's not too farfetched to assume they'll take 3-4 hours and a few hundred dollars of tokens to find a target-specific zero-day.

4. VULNERABILITIES: Thousands of Leaked AWS Access Keys Still Active
5. CYBERCRIME: Unmasking TeamPCP: Software Supply Chain Attacks – Flare

   These younger criminals are finding out that it's hard to not get caught when you're terminally online...

6. JAILBREAKS: Researcher shows how Claude Code can be tricked simply by asking it to summarize a website
7. ESSAYS: Security Research Without Asking Permission

   Neils Provos discusses his experiences kicking paid services to the curb and using open models on his own hardware to do security work. TL;DR - you can do a lot of interesting work with $10k of hardware.

8. DATA LEAKS: FBI Probes Service Selling 153M+ Drivers Licenses – Krebs on Security

   This one is really bad, as photos of all the IDs were included in the breach. My thoughts:

   1. The attackers took down the site, and I'm not sure that's a good thing. It could suggest that the criminals were worried about getting too much attention from US law enforcement, OR it could suggest that one customer made an offer they couldn't refuse to sell the whole database, which required them to take it down. We've often seen criminals willing to both sell individual data records, or all the stolen data wholesale.
   2. This much data would have taken a long time to steal - there was definitely a detection failure somewhere.
   3. We don't know for sure that it was IDScan yet - I'm sure we'll learn more in the next 24-48 hours.
   4. If it was IDScan, this was definitely a 'you had one job' situation. Any vendor knowingly capturing millions of ID cards must assume they're painting a target on their back and should be spending a lot more resources on security compared to the average business.
   5. The majority of these IDs include photos. Reverse image search engines exist, that allow you to search the Internet with someone's face. With hundreds of millions of ID photos, it's possible to deanonymize people in witness protection programs, people who have changed their names or identities for any reason, discover people using fake ids, identify people with medical IDs for marijuana and other controlled substances, determine where people work (CAC cards and commercial driver licenses).
9. SQUIRREL: Microduck – A tiny biped robot you can teach new tricks
10. SQUIRREL: Dyson CameraJet™ electric toothbrush (Ceramic pink)
11. BREACHES: Hackers push malicious Virtualizor update in BGP hijacking attack