Loading story…
Recent research highlights the growing security risks posed by advances in artificial intelligence, particularly in the areas of large language model (LLM) steganography and hypergame theory for AI training. LLM-based steganography enables the embedding of undetectable secret payloads within any text, across all media formats, leveraging the ubiquity of AI systems and making detection nearly impossible. This technique represents a fundamental shift from traditional covert communication methods, as it requires no specialized equipment and can operate at a global scale, raising concerns about the integrity of text-based infrastructure and the potential for widespread, undetected data exfiltration.
In parallel, the application of hypergame theory to high-stakes AI training is revolutionizing how adversarial scenarios are modeled and understood. New frameworks allow for the analysis of strategic misalignment in multi-agent systems, addressing the reality that adversaries often operate with different mental models and exploit hidden assumptions. These advances are being integrated into military and cybersecurity operations, enhancing the realism and effectiveness of AI-driven defense strategies. Together, these developments underscore the urgent need for new detection, defense, and governance mechanisms as AI capabilities outpace traditional security paradigms.
Track how attackers are adapting to this technology.
2 events from the most recent confirmed update back to the earliest known activity.
A paper by Vince Trencsenyi of Royal Holloway University presented an automated, logic-based framework for analyzing strategic misalignment in multi-agent systems using hypergame rationalization. The work defines mechanisms such as Strong and Weak Hypergame Nash Equilibria and supports post-incident analysis and AI training.
Research by Antonio Norelli and Michael Bronstein showed that large language models can hide a full secret payload inside an equally long, innocent-looking text. The write-up says the method can embed multiple hidden messages and is not detectable by current forensic or statistical techniques.
Follow how adversaries are adapting to this technology, and where it touches your stack today.
2 references tracked. Mallory keeps watching after this page renders.
Map indicators from this story to your assets and identify affected systems in minutes.
Every observed campaign, victim, and pivot linked to actors named in this story.
Malware, exploits, and IOCs connected to the activity described here.
YARA, Sigma, and Snort rules deployed to your SIEM as soon as they’re published.
Get matching new stories delivered to your team as they break — not the next morning.
Ask questions about this story and take action on the answers.
Other stories Mallory is tracking that overlap on entities, sources, or topic labels.
Recent research and expert commentary highlight the growing sophistication of attacks targeting artificial intelligence systems, particularly large language models (LLMs) and multimodal AI. Studies reveal that attackers can exploit weaknesses in open weight LLMs through multi-turn prompt injection, significantly increasing the success rate of bypassing safety guardrails compared to single-turn attacks. These findings underscore the importance of robust alignment strategies and continuous adversarial testing, as attackers adapt their tactics over extended interactions to elicit harmful or unauthorized outputs from AI models. In addition to text-based prompt injection, multimodal AI systems face unique threats from hidden or invisible instructions embedded in both text and visual data. Techniques such as white-on-white text in resumes or hidden prompts in online profiles can manipulate AI-driven screening tools, allowing malicious actors to influence outcomes undetected by human reviewers. These developments demonstrate that as AI systems become more integrated into critical decision-making processes, their attack surface expands, necessitating new approaches to securing both the models and their input channels against increasingly subtle and complex adversarial techniques.
Jun 29, 2026
Cybersecurity threats leveraging artificial intelligence are rapidly evolving, with attackers using AI to automate vulnerability discovery, conduct deepfake-enabled social engineering, and compromise large language models (LLMs) through data poisoning and malicious model files. A Gartner survey found that 62% of organizations experienced deepfake attacks, often combined with social engineering to defraud companies, while 32% faced attacks on AI applications such as prompt injection. Security researchers have demonstrated prompt injection vulnerabilities in AI summarization tools, highlighting the risks of insufficient separation between user input and system prompts. Additionally, experts warn that LLMs can be compromised via poisoned training data, malicious adapters, or trojanized model files, expanding the attack surface for organizations adopting AI technologies. These developments underscore the need for robust AI governance, technical defenses, and employee awareness to counter increasingly sophisticated AI-driven threats.
Jun 29, 2026
The rapid integration of artificial intelligence (AI) and large language models (LLMs) into cybersecurity operations and software development is fundamentally altering both the attack surface and defensive strategies. Security teams are leveraging AI to automate alert triage, summarize threat intelligence, and streamline incident response, while organizations like Microsoft are bundling AI-powered security assistants such as Security Copilot with enterprise products to democratize advanced threat detection and response. However, this shift introduces new risks, including prompt injection attacks, the challenge of validating AI-generated code, and the emergence of "vibe coding," where natural language prompts replace traditional software engineering rigor, potentially leading to insecure or unmaintainable code. Studies show that while LLMs can assist in patching known vulnerabilities, their effectiveness drops with unfamiliar or artificially altered code, highlighting limitations in current AI capabilities for secure software maintenance. The evolving AI attack surface is characterized by probabilistic model behavior, making vulnerabilities less predictable and harder to patch compared to traditional software flaws. Security experts warn that the speed and scale enabled by AI can benefit both defenders and attackers, with concerns about AI-enabled autonomous attacks and the need for new security models to address reasoning manipulation rather than just input validation. As organizations increase cybersecurity budgets and invest in AI-driven solutions, the industry faces a dual imperative: harnessing AI's potential to improve defense while developing robust controls and validation processes to mitigate the novel risks it introduces.
Jun 29, 2026