Cyber Threat Weekly – #89
The week of July 28th through August 3rd, around 386 cyber news articles were reviewed. A moderate amount of cyber threat trends and adversarial behavior news to share. Been thinkin about AI and its self-preservation tendencies.
In an experiment, multiple models chose blackmail the majority of the time to save themselves. An Agentic AI deleted a production database during a code-freeze, then was deceptive about what it had done. Researchers building the large language models are worried that AI may try to hide its chain-of-thought so humans can’t see what it’s doing. We may want to slow down a little and ensure safeguards are in place.
Let’s start with 2025 Incident Response Report: Social Engineering Edition. SonicWall firewall devices targeted in ransomware attacks. Researchers share threat actor groups they track. Cursor IDE prompt-injection attack details.
Open-source repository attack stories. With all the buzz of agentic AI, here is a bit of a buzz kill. Ransomware gangs ratcheting up physical threats, new report. 2025 Midyear Global Threat Intelligence Index, infostealers are killing it.
Another report for correlation, ThreatLabz 2025 Ransomware Report. Researchers test LLMs in autonomous cyberattacks.
Broken Record Alert: Don’t get pwned by N-day vulnerabilities!!!
Known exploited software flaws are one of the top 4 initial access vectors and have increased sharply in recent months. We continue to share n-day vulnerabilities being actively exploited. Priority #1, start with the CISA / VulnCheck known exploited vulnerability (KEV) catalogs. If it’s in the catalog, it should be patched.
A close #2 priority is flaws with weaponized proof of concept (PoC) code available. Exploit chances are higher with weaponized PoC code available. If you do nothing else with patching, have an emergency 24-to-48-hour patching process for actively exploited and weaponized PoC code available vulnerabilities.
You should consider what is exposed to the Internet. Architecture and zero trust network access (ZTNA) can go a long way to minimizing the number of devices and services exposed to the Internet.
CISA Known Exploited Vulnerabilities – July 28th to August 3rd:
CVE-2023-2533 – PaperCut NG/MF Cross-Site Request Forgery (CSRF) Vulnerability:
Under specific conditions, could potentially enable an attacker to alter security settings or execute arbitrary code.
CVE-2025-20337 – Cisco Identity Services Engine Injection Vulnerability:
Due to insufficient validation of user-supplied input, an attacker could exploit this vulnerability by submitting a crafted API request. Successful exploitation could allow an attacker to perform remote code execution and obtaining root privileges on an affected device.
CVE-2025-20281 – Cisco Identity Services Engine Injection Vulnerability:
Due to insufficient validation of user-supplied input, an attacker could exploit this vulnerability by submitting a crafted API request. Successful exploitation could allow an attacker to perform remote code execution and obtaining root privileges on an affected device.
2025 Incident Response Report: Social Engineering Edition
This one starts out with our biggest problem set, right out the gate. After laying out five key social engineering tends, three systematic enablers surface. The enablers are over-permissioned access, gaps in behavioral visibility, and unverified user trust in human processes. The numbers are shared as well as defensive recommendations.
https://cyberscoop.com/social-engineering-top-attack-vector-unit-42/
Akira Ransomware Targeting SonicWall Firewalls
It appears a possible zero-day bug is being abused to compromise SMA 100 appliances. Researchers observe rootkit malware install and cleaning of logs to hinder investigations. This one is stealthy, hunting guidance provided.
Unit 42 Shares Tracked Threat Actor Groups
While not behavioral, the list of threat actors and also known as (aka) names maybe useful for threat intelligence folks attempting to keep up with the various threat actor names. Also, a quick summary and targeted verticals is shared.
https://unit42.paloaltonetworks.com/threat-actor-groups-tracked-by-palo-alto-networks-unit-42/
Prompt-injection Attack in AI-powered Cursor IDE
This is more for tracking purposes, we’ll be seeing a lot more of these kinds of attacks. Tracked as CVE-2025-54123 and dubbed CurXecute. The Cursor IDE has support for Model Context Protocol (MCP) allowing connections to external resources.
https://www.aim.security/lp/aim-labs-curxecute-blogpost
Continued Targeting of Open-Source Repositories
This is the collection of attack stories around npm packages, PyPI, and more. Interestingly, the first story includes an AI generated malicious npm package.
https://thehackernews.com/2025/08/ai-generated-malicious-npm-package.html
https://getsafety.com/blog-posts/threat-actor-uses-ai-to-create-a-better-crypto-wallet-drainer
https://therecord.media/north-korean-hackers-targeting-open-source-repositories
https://thehackernews.com/2025/07/pypi-warns-of-ongoing-phishing-campaign.html
https://thehackernews.com/2025/07/hackers-breach-toptal-github-publish-10.html
Surprise, Agentic AI and Vibe Coding aren’t Ready for Prime Time
We are going to continue to hear of attacks on agentic AI, it’s simply too new. Technology deployment always outpaces security maturity around that technology. Will we get there, yes, but you should start slow, focus on specific use cases, and always have a human in the loop. Vibe coding is having similar growing pains.
https://www.lasso.security/blog/identitymesh-exploiting-agentic-ai
https://www.darkreading.com/application-security/cybersecurity-vibe-check-vibe-coding
https://www.veracode.com/wp-content/uploads/2025_GenAI_Code_Security_Report_Final.pdf
2025 Ransomware Risk Report
A survey of 1,500 IT and security professional sheds some light on the ransomware menace. Physical threats were leveraged in 40% of attacks. Identity was a huge attack vector, 83%. Companies successfully ransomed, 73% were attacked multiple times.
https://www.infosecurity-magazine.com/news/ransomware-attacks-escalate/
2025 Midyear Global Threat Intelligence Index
Credential theft from infostealers up 800%, solidifying identity as a prolific attack vector. Ransomware continues to be one of the top financial risks to business operations, incidents are up 179% for H1 2025. Data breaches are up 235%.
ThreatLabz 2025 Ransomware Report
This report spans April 2024 to April 2025. Like other reports ransomware attempts blocked by Zscaler were up 145% year-over-year. Ransomware attacks surge world-wide. The United States continues to be targeted with 50.8% of attacks.
https://www.cybersecuritydive.com/news/zscaler-ransomware-report-manufacturing-targeted/756147/
https://www.zscaler.com/resources/industry-reports/threatlabz-2025-ransomware-analysis.pdf
LLMs Autonomously Carry Out Sophisticated Cyberattacks
Researchers set a goal to measure a large language models’ ability to plan and carry out a cyberattack. The research found that LLMs, when provided with an abstracted ‘mental model’ of red teaming behavior and actions, can plan and initiate autonomous attacks.
https://www.cybersecuritydive.com/news/research-llms-attacks-without-humans/754203/
https://engineering.cmu.edu/news-events/news/2025/07/24-when-llms-autonomously-attack.html
31337 InfoSec - Cyber Threat Weekly - Derek Krein Newsletter
Join the newsletter to receive the latest updates in your inbox.
Comments
Sign in to join the conversation.
Just enter your email below to receive a login link.
