New attack bypasses AI guardrails by encrypting malicious prompts

SCM feed for Latest

Overview

Researchers have identified a new method called 'cryptographic context injection' that allows attackers to bypass security measures in AI systems. Developed by Rony Utevsky from Adversa, this attack takes advantage of how AI models interpret data, specifically by encrypting malicious prompts. This is significant because it could enable harmful instructions to be processed by AI without detection, potentially leading to misuse in various applications. As AI continues to be integrated into more systems, understanding and addressing these vulnerabilities is crucial for maintaining security and trust in AI technologies. Companies that rely on AI for decision-making or automation should be particularly vigilant about this emerging risk.

Key Takeaways

  • Affected Systems: AI models and systems that process user inputs.
  • Action Required: Implement stronger validation and monitoring of input prompts to detect and mitigate encrypted malicious prompts.
  • Timeline: Newly disclosed

Original Article Summary

The attack, developed by Rony Utevsky, a researcher at Adversa, is dubbed "cryptographic context injection" and exploits a fundamental weakness in how AI models process information.

Impact

AI models and systems that process user inputs.

Exploitation Status

The exploitation status is currently unknown. Monitor vendor advisories and security bulletins for updates.

Timeline

Newly disclosed

Remediation

Implement stronger validation and monitoring of input prompts to detect and mitigate encrypted malicious prompts.

Additional Information

This threat intelligence is aggregated from trusted cybersecurity sources. For the most up-to-date information, technical details, and official vendor guidance, please refer to the original article linked below.

Related Topics: This incident relates to Vulnerability.

Related Coverage

AI supply chain risk is showing up in developer workflows first

Help Net Security

In a recent interview, Dr. Jaushin Lee, CEO of Zentera Systems, pointed out that AI supply chain risks are primarily affecting developer workflows and open-source package repositories. He noted that while some issues like poisoned model weights and compromised servers are mainly seen in research settings, the real-world impact is felt in the everyday work of developers. Dr. Lee emphasized that companies might gain more risk reduction from proper segmentation of their systems than from investing heavily in new tools. He also mentioned the shortcomings of self-hosting AI models and suggested that software teams could benefit by adopting certain semiconductor isolation practices. This conversation brings attention to the evolving risks in AI development and the need for better security practices in software development environments.

Aug 25, 2026

New TCG guidance gives buyers a way to test PQC-ready TPM claims

Help Net Security

The Trusted Computing Group (TCG) has released new guidelines for Trusted Platform Modules (TPMs), which are essential components that secure a device's cryptographic keys and firmware integrity. These guidelines specifically address how TPMs can be deemed quantum-safe, a growing concern as quantum computing technology advances. Now, buyers can request proof from vendors that their TPMs meet these new standards, ensuring they are prepared for future quantum threats. This is significant for companies looking to protect their data from potential quantum attacks, as it provides a way to verify the security claims made by manufacturers. The move aims to enhance the overall security landscape as businesses transition to quantum-resistant technology.

Aug 25, 2026

ShinyHunters claims social engineering attack against ReliaQuest

SCM feed for Latest

A group known as ShinyHunters claims to have executed a social engineering attack against ReliaQuest. The attackers targeted employees by calling them and attempting to deceive them into visiting a fraudulent single sign-on (SSO) page. This fake page was hosted on a lookalike domain, reliaquest[.]claims, designed to mimic the legitimate ReliaQuest site. Such tactics can lead to credential theft and unauthorized access to sensitive company data. This incident raises concerns about the effectiveness of security training and awareness among employees, as social engineering remains a prevalent threat in cybersecurity.

Aug 24, 2026

WordPress plugin vulnerabilities allow admin account takeover

SCM feed for Latest

Researchers have identified two vulnerabilities in WordPress plugins, tracked as CVE-2026-61979 and CVE-2026-15981, that can be exploited together to bypass authentication and potentially take over admin accounts. This poses a significant risk to users of affected plugins, as attackers could gain unauthorized access to sensitive areas of WordPress sites. The vulnerabilities are particularly concerning for website administrators who may not be aware of these security flaws. It's crucial for users to check if their plugins are affected and take appropriate action to secure their sites, especially since the potential for exploitation exists. Prompt updates and vigilance are key to maintaining site security in light of these findings.

Aug 24, 2026

Developer alleges Alibaba uses audio fingerprinting for web tracking

SCM feed for Latest

Matt Callaghan, a software engineer, has raised concerns that Alibaba's website may be using audio fingerprinting techniques for tracking users. He found that the site employs obfuscated audio scripts that create a waveform and analyze its output, which could allow the company to monitor user behavior in a way that bypasses traditional tracking methods. This discovery raises significant privacy issues, as it suggests that users may be unwittingly tracked through audio signals emitted from their devices. The implications are serious, especially for individuals who value their privacy online. The use of such techniques could lead to increased scrutiny from regulators and may prompt users to reconsider their interactions with the site.

Aug 24, 2026

Exploited Zimbra Flaw Highlights Shrinking Window to Patch

darkreading

The Cybersecurity and Infrastructure Security Agency (CISA) has issued an urgent three-day deadline for agencies to address a serious vulnerability in Zimbra, identified as CVE-2026-73570. This flaw enables attackers to take complete control over a user's communications, posing a significant risk to organizations using this software. The vulnerability could lead to unauthorized access to sensitive information and disrupt business operations. As Zimbra is widely used for email and collaboration, the implications of this vulnerability are considerable, affecting both public and private sector entities. Agencies are urged to act quickly to implement the necessary patches to mitigate this risk.

Aug 24, 2026