As the 2023 Executive Order on Artificial Intelligence (AI) specifically lays out, “Harnessing AI for good and realizing its myriad benefits requires mitigating its substantial risks.” One significant but often overlooked threat in AI data is steganography—the art of hiding malicious or sensitive data within seemingly benign files, including images or videos. As agencies increasingly leverage AI for mission-critical tasks, protecting against hidden threats like steganography becomes paramount.

Why is Steganography a Critical Risk to AI in Defense?

AI systems are more interconnected and data-driven than ever, making them ripe targets for steganographic attacks. According to a recent NIST report, adversarial machine learning poses grave dangers to AI infrastructure. Attackers can corrupt training datasets with malicious or poisoned data, causing AI models to make incorrect or dangerous decisions. These adversarial manipulations include:

• Evasion attacks – altering an input to trick the AI post-deployment
• Poisoning attacks – corrupting data at the training stage
• Prompt injection – text prompts designed to enable the user to perform unintended or unauthorized actions
• Privacy attacks – extracting or reconstructing sensitive information from training or input data

Moreover, AI-based steganography has recently taken an even more insidious turn. A quirk in the Unicode standard allows for the embedding of invisible characters that AI models, like GPT-4 or Claude, can read but are imperceptible to human users. Such methods, labeled as “ASCII smuggling,” are used to exfiltrate sensitive information covertly from platforms like Microsoft Copilot.

Other Emerging AI Exploitation Threats

Beyond inserting malicious or poisoned data, prompt injection and adversarial machine learning represent evolving cyber risks to AI. As exemplified by proof-of-concept (POC) attacks on Microsoft Copilot, researchers show how hidden characters can direct AI models to leak confidential data. The NIST report also elaborates on the various types of AI-focused attacks, such as evasion, privacy, and abuse attacks. These threats are not just theoretical, with evidence of their use already appearing in the field.

How does Owl Cyber Defense Protect AI Systems?

Owl Cyber Defense’s advanced cross domain solutions (CDS) are purpose-built to detect and mitigate data threats transferred between sensitive systems, including steganographic content embedded within data. Using advanced data and provenance validation, encode/decode capabilities, and sophisticated content filtering including destructive filtering, Owl secure transfer solutions provide comprehensive protection for U.S. civilian and defense organizations adopting AI.

1. Threat Detection & Data Provenance – Owl CDSs employ sophisticated filtering technology which filters audio & video media for potential embedded data. Metadata tags, security/classification markings, and data provenance are also vetted.
2. Steganographic Mitigation Techniques – Owl CDSs utilize various sophisticated filtering technologies to remove hidden or embedded steganographic information. These techniques may include transcoding the media, applying destructive filtering, and other methods that filter or compromise the steganographic data without affecting the quality of the original data stream.
3. Logging & Quarantine – Upon identifying potential steganographic content, Owl CDSs log incidents and quarantine data that does not conform to security policy. Data payload in suspect signaling data can also be extracted and quarantined.

Empowering Agencies to Safely Utilize AI

AI offers unparalleled benefits for defense operations, from rapid decision-making to enhanced situational awareness. However, these advantages can only be fully realized if the data feeding AI systems is secure. Owl Cyber Defense’s solutions empower agencies to adopt AI confidently, knowing that their systems are protected from covert threats like steganography.

By integrating Owl’s cross domain solutions into existing AI frameworks, defense agencies can ensure the integrity of their data and maintain a strong security posture. This proactive approach not only secures AI systems but also enables agencies to fully leverage AI’s transformative potential.

The Future of AI in Defense: Protecting Against Hidden Threats

The rapid adoption of AI in defense is undeniable, but so are the threats that come with it. Protecting AI systems from steganography is a crucial step in ensuring mission success. Owl Cyber Defense offers the cutting-edge technology needed to detect and mitigate these hidden threats, safeguarding the future of AI-driven operations.

Want to learn more about how Owl Cyber Defense can enhance the security of your AI systems?
Download our free white paper:
Enabling & Securing AI Utilization in Defense with Cross Domain Solutions