Security
Agentic ProbLLMs: Exploiting AI Computer-Use and Coding Agents
This talk demonstrates end-to-end prompt injection exploits that compromise agentic systems. Specifically, we will discuss exploits that target computer-use and coding agents, such as Anthropic's Claude Code, GitHub Copilot, Google Jules, Devin AI, ChatGPT Operator, Amazon Q, AWS Kiro, and others.
Exploits will impact confidentiality, system integrity, and the future of AI-driven automation, including remote code execution, exfiltration of sensitive information such as access tokens, and even joining Agents to traditional command and control infrastructure. Which are known as "ZombAIs", a term first coined by the presenter as well as long-term prompt injection persistence in AI coding agents.
Additionally, we will explore how nation state TTPs such as ClickFix apply to Computer-Use systems and how they can trick AI systems and lead to full system compromise (AI ClickFix).
Finally, we will cover current mitigation strategies and forward-looking recommendations and strategic thoughts.
During the Month of AI Bugs (August 2025), I responsibly disclosed over two dozen security vulnerabilities across all major agentic AI coding assistants. This talk distills the most severe findings and patterns observed.
Key highlights include:
* Critical prompt-injection exploits enabling zero-click data exfiltration and arbitrary remote code execution across multiple platforms and vendor products
* Recurring systemic flaws such as over-reliance on LLM behavior for trust decisions, inadequate sandboxing of tools, and weak user-in-the-loop controls.
* How I leveraged AI to find some of these vulnerabilities quickly
* The AI Kill Chain: prompt injection, confused deputy behavior, and automatic tool invocation
* Adaptation of nation-state TTPs (e.g., ClickFix) into AI ClickFix techniques that can fully compromise computer-use systems.
* Insights about vendor responses: from quick patches and CVEs to months of silence, or quiet patching
* AgentHopper will highlight how these vulnerabilities combined could have led to an AI Virus
Finally, the session presents practical mitigations and forward-looking strategies to reduce the growing attack surface of probabilistic, autonomous AI systems.
Additional information
| Live Stream | https://streaming.media.ccc.de/39c3/zero |
|---|---|
| Type | Talk |
| Language | English |
More sessions
| 12/27/25 |
The Great Firewall of China (GFW) is one of, if not arguably the most advanced Internet censorship systems in the world. Because repressive governments generally do not simply publish their censorship rules, the task of determining exactly what is and isn’t allowed falls upon the censorship measurement community, who run experiments over censored networks. In this talk, we’ll discuss two ways censorship measurement has evolved from passive experimentation to active attacks against the Great ...
|
| 12/27/25 |
Reports of GNSS interference in the Baltic Sea have become almost routine — airplanes losing GPS, ships drifting off course, and timing systems failing. But what happens when a group of engineers decides to build a navigation system that simply *doesn’t care* about the jammer? Since 2017, we’ve been developing **R-Mode**, a terrestrial navigation system that uses existing radio beacons and maritime infrastructure to provide independent positioning — no satellites needed. In this talk, ...
|
| 12/27/25 |
Zwei Jahre nach dem ersten KIM-Vortrag auf dem 37C3: Die gezeigten Schwachstellen wurden inzwischen geschlossen. Weiterhin können mit dem aktuellen KIM 1.5+ nun große Dateien bis 500 MB übertragen werden, das Signaturhandling wurde für die Nutzenden vereinfacht, indem die Detailinformationen der Signatur nicht mehr einsehbar sind. Aber ist das System jetzt sicher oder gibt es neue Probleme?
|
| 12/27/25 |
While trying to apply fault injection to the AMD Platform Security Processor with unusual (self-imposed) requirements/restrictions, it were software bugs which stopped initial glitching attempts. Once discovered, the software bug was used as an entry to explore the target, which in turn lead to uncovering (and exploiting) more and more bugs, ending up in EL3 of the most secure core on the chip. This talk is about the story of trying to glitch the AMD Platform Security Processor, then ...
|
| 12/27/25 |
The Deutschlandticket was the flagship transport policy of the last government, rolled out in an impressive timescale for a political project; but this speed came with a cost - a system ripe for fraud at an industrial scale. German public transport is famously decentralised, with thousands of individual companies involved in ticketing and operations. Unifying all of these under one national, secure, system has proven a challenge too far for politicians. The end result: losses in the hundreds of ...
|
| 12/27/25 |
In August 2024, Raspberry Pi released their newest MCU: The RP2350. Alongside the chip, they also released the RP2350 Hacking Challenge: A public call to break the secure boot implementation of the RP2350. This challenge concluded in January 2025 and led to five exciting attacks discovered by different individuals. In this talk, we will provide a technical deep dive in the RP2350 security architecture and highlight the different attacks. Afterwards, we talk about two of the breaks in ...
|
| 12/27/25 |
FreeBSD’s jail mechanism promises strong isolation—but how strong is it really? In this talk, we explore what it takes to escape a compromised FreeBSD jail by auditing the kernel’s attack surface, identifying dozens of vulnerabilities across exposed subsystems, and developing practical proof-of-concept exploits. We’ll share our findings, demo some real escapes, and discuss what they reveal about the challenges of maintaining robust OS isolation.
|
