Downloads: 2
Research Paper | Computer Engineering | Volume 15 Issue 2, February 2026 | Pages: 988 - 991 | India
Adversarial Attacks on Agentic AI Systems: Mechanisms, Impacts, and Defense Strategies
Abstract: Agentic AI is starting to show up more, and its kind of moving away from just those big language models that sit there and answer questions to something more active, like systems that can go after goals on their own and deal with all sorts of digital stuff or even real world things. But with that freedom comes a lot more ways for attacks to happen, and the old ways of securing just the model itself do not cut it anymore. I think the main issue here is how these agents have these loops where they reason step by step, keep memories around, and get access to more things, which opens up new weak spots. This paper looks into that threat side pretty carefully, pulling from design science research and going over more than 160 recent studies to sort out the attacks into different parts, like the model layer, tools they use, memory stuff, and how they all coordinate. In the experiments, they used this BAD-ACTS thing to test, and it turns out adversarial agents can mess with the system a lot, especially when theres database access or big decisions involved, with really high success rates. That part stands out because it shows how behaviors get twisted in those setups. One big thing they point out is this lethal trifecta, where access to sensitive data mixes with exposure to untrusted content and talking to outside systems, and that drives a ton of the overall risk. It feels like without handling that combo, the whole system could fall apart. To fight back, theres this multilayered defense called MAAIS, and then the 4C framework, which covers core elements, connections, how they think, and compliance rules, aiming for better resilience instead of just stopping attacks one by one. Some people might say its still early to tell if these work perfectly in real scenarios. Overall, for anyone building or regulating this AI stuff, these ideas seem key to making autonomous systems that people can trust in a world full of bad actors, though I am not totally sure how fast that shift will happen.
Keywords: Agentic AI Security, Adversarial Attacks, Prompt Injection, Autonomous Systems, Threat Modeling, MAAIS Framework
How to Cite?: Syeda Ameena, Akheel Mohammed, "Adversarial Attacks on Agentic AI Systems: Mechanisms, Impacts, and Defense Strategies", Volume 15 Issue 2, February 2026, International Journal of Science and Research (IJSR), Pages: 988-991, https://www.ijsr.net/getabstract.php?paperid=SR26216005631, DOI: https://dx.dx.doi.org/10.21275/SR26216005631