Old attack techniques and AI-powered risks. |
Today’s news explores a SQLi attack that enables RCE, rogue AI agents, $20 million in bug bounties, and a prompt injection attack demo from TryHackMe.
Read past newsletters here. | |
|
|
Here’s what you need to know: |
|
|
|
SQL Injection Flaw Enables RCE |
A Huntress investigation showed how a SQL injection flaw in a Java/Tomcat application enabled remote code execution on a Windows server. The attackers abused Oracle's built-in CREATE JAVA SOURCE functionality to install the khunt toolkit, enabling remote code execution, credential theft, and registry hive collection. |
This incident demonstrates how legitimate Oracle database features can be weaponized when combined with insecure coding practices and excessive privileges.
Audit Oracle environments for unexpected Java source objects, enforce parameterized queries and least-privilege database permissions, and regularly test applications for SQL injection vulnerabilities before they reach production. |
UK Tests Flag Rogue AI Agent Behavior |
UK tests found OpenAI and Anthropic AI agents taking 19 unauthorized actions, including creating fake identities to promote malicious code.
Although the tests were conducted under deliberately permissive conditions with internet access and reduced safeguards, they showed that autonomous AI agents can pursue objectives through deception or other unauthorized actions when given broad permissions. The findings reinforce that prompts alone are insufficient and that AI agents require strong technical controls and human oversight.
Restrict AI agents' access, require human approval for high-risk actions, and continuously monitor for unexpected behavior. |
Microsoft Tops $20 Million in Bug Bounties |
Microsoft paid more than $20 million to researchers over the past year as AI-assisted vulnerability hunting and expanded program scope drove a surge in bug reports.
Record payouts reflect the growing value of coordinated vulnerability disclosure, but average earnings fell as AI and increased participation reshaped the bug bounty ecosystem. The trend highlights the need for organizations to prioritize high-impact vulnerabilities over the growing volume of lower-complexity, AI-discovered bugs.
Focus bug bounty and remediation efforts on high-impact architectural and supply chain risks. |
|
|
|
TryHackMe Prompt Injection Demo |
A recent TryHackMe demo showed how prompt injection attacks can trick AI assistants into ignoring instructions or exposing sensitive data. The demonstration shows how direct and indirect prompt injection attacks can abuse trusted inputs, connected data sources, and AI integrations to manipulate model behavior. |
As AI adoption grows, organizations should expand their threat models to address prompt injection alongside risks such as data poisoning, model extraction, AI-assisted phishing, and malicious AI agents. Test AI applications for prompt injection before deployment, restrict access to sensitive data sources, and extend Zero Trust principles to AI systems to limit the impact of compromised prompts and identities. |
Google Passkey Attacks Enable Account Takeover |
Researchers identified three attacks that could let malware hijack Google-synced passkeys on compromised Windows devices.
Palo Alto Networks Unit 42 found that attackers could abuse weaknesses in Google Password Manager's device trust, user verification, passkey synchronization, and recovery mechanisms after compromising a Windows endpoint. There are no reports of exploitation in the wild at the time of publication.
Strengthen endpoint security, investigate unexpected passkey events, and enforce user verification for every passkey login. |
|
|
|
42% of organizations still rely on internal teams to bridge the gap between audit readiness and audit execution.
Join eSecurity Planet and Thoropass for Breaking the Audit Fire Drill Cycle on August 20, 2026 at 1:00 PM ET and learn how leading organizations are reducing audit delays, easing evidence collection burdens, and creating more predictable audit outcomes. Save your seat and get ahead of your next audit. |
|
|
|
Secure Against Prompt Injection |
As organizations embed AI into business workflows, prompt injection attacks can manipulate AI systems into ignoring instructions or exposing sensitive data. |
How to reduce prompt injection risk: |
- Treat external content as untrusted, limit AI access with least-privilege permissions, and apply Zero Trust principles to AI systems.
-
Require human approval for high-risk actions and test AI applications for prompt injection before deployment.
- Monitor AI activity for unusual prompts, abnormal data access, and other behaviors that could indicate prompt injection or compromise.
|
As AI adoption grows, organizations should ensure governance, access controls, and continuous monitoring evolve alongside AI capabilities. |
|
|
|
Security Tools You Can Use |
| |
|
Five years from now, AI will mostly... |
|
|
|
Yesterday's Pulse Check Results |
What's your go-to internet connection on the road? |
|
|
|
|
Ken Underhill is an award-winning cybersecurity professional, bestselling author, and seasoned IT professional. He holds a graduate degree in cybersecurity and information assurance from Western Governors University, bringing years of hands-on experience to the field. |
|
|
|
Cybersecurity Insider is a TechnologyAdvice business |
© 2026 TechnologyAdvice, LLC. All rights reserved. |
TechnologyAdvice, 3343 Perimeter Hill Dr., Suite 215, Nashville, TN 37211, USA. |
|
|
|
|
