A concise, fact-based update for security and risk professionals covering the past 24 hours.

🔐 Core Security Intelligence

1) HashJack: new indirect prompt-injection technique targeting AI browsers

What’s new:
Researchers at Cato CTRL disclosed a new attack method called HashJack. It works by embedding malicious instruction fragments after the “#” in URLs, content that AI-enabled browsers send to their AI assistant even though traditional security layers ignore fragments.

Source: HashJack: A New Attack That Fools AI Browsers

Why it matters:
AI-enabled browsers merge browsing with automated reasoning. Because URL fragments never reach the server, existing network-security tools don’t detect them. That means attackers can weaponize normal websites to inject instructions directly into a user’s AI assistant, bypassing conventional defenses entirely.

Defenses:

• Strip or sanitize URL fragments before AI processing. Browser-integrated AI should treat all fragment content as untrusted.
• Isolate AI browser sessions. Use sandboxes or separate profiles for AI-augmented browsing sessions.
• Log AI-browser interactions. Monitor for unexpected API calls, external requests, or automations triggered solely from browsing activity.

Expert Insight:
AI browsers represent the next major attack surface. Traditional URL inspection is no longer enough when malicious instructions hide in client-side-only fragments that AI logic still interprets.

2) Criminal underground markets now selling custom malicious LLMs

What’s new:
CyberScoop reports a growing black market for modified or fully malicious LLMs designed for exploitation: automated recon, exploit creation, phishing orchestration and data exfiltration. These tools mimic the capabilities of WormGPT and similar models but are more advanced and tailored for cybercrime.

Source: Malicious LLM Tools Sold on Criminal Underground Markets

Why it matters:
Attackers no longer need high technical skill, LLM-as-a-hacker tools provide them with exploit logic, vulnerability insights and code-generation capabilities. This accelerates attack volume and sophistication, democratizing cybercrime.

Defenses:

• Enhance AI-aware threat intelligence. Expect attackers to use LLMs for recon, exploit assistance, and scalable phishing.
• Monitor for AI-automation patterns. Watch for unusual bursts of scripted activity, rapid code generation, or repeated external model calls.
• Require strict change-control in dev environments. Mandate human review for all code introduced, especially if AI-assisted.

Expert Insight:
We’re entering a phase where “skill” is no longer a barrier to cybercrime. Automated exploitation will increase attack density; defenders must match automation with automation.

3) Olymp Loader: new “fully undetectable” Malware-as-a-Service

What’s new:
A new MaaS offering called Olymp Loader is being advertised on cybercrime forums. It bundles crypter, dropper and info-stealer capabilities into a single FUD (Fully Undetectable) loader, providing advanced evasion even for low-skilled attackers.

Source: Hackers Advertise New Olymp Loader MaaS Offering

Why it matters:
The sophistication once associated with APT-grade loaders is now fully commoditized. This increases background threat noise and overwhelms organizations relying on signature-based detection.

Defenses:

• Adopt behavioral EDR controls. Detect process hollowing, injection sequences, and obfuscated memory activity, not just signatures.
• Apply layered segmentation. Limit lateral movement and reduce blast radius even if a loader executes locally.
• Harden endpoint privilege. Ensure least privilege, exploit-mitigation policies and application-control baselines.

Expert Insight:
Threat volume is rising because attackers can buy sophisticated tools instead of developing them. This requires defenders to optimize detection for scale, not rarity.

⚠️ Updates / Follow-ups

No verified updates were published in the last 24 hours for previously covered briefings.

Summary Table

Categories: Cybersecurity News

Tags: AI Security, LLM Threats, Malware-as-a-Service, Prompt Injection