New study from George Mason University claims AI can be hacked with a simple 'typo' in its memory

Researchers at George Mason University have demonstrated that modern AI systems can be covertly compromised by flipping a single bit in the model’s stored parameters, a change the team compares to a tiny “typo” in memory. Their attack, dubbed OneFlip, uses hardware-based fault techniques similar to Rowhammer to target specific memory cells holding neural network weights, altering one 0/1 value to embed a backdoor without needing to retrain the model, modify its code, or visibly reduce its overall accuracy. According to the study, a successfully executed OneFlip attack can cause deep learning models to behave normally under almost all conditions but produce attacker-controlled wrong outputs when a particular trigger input appears. The researchers highlight risks for safety-critical applications such as self‑driving cars, where this could cause misclassification of a stop sign, or facial recognition systems, where a targeted person could be misidentified while the system otherwise passes standard tests. Because the manipulation occurs at the hardware/memory level, it may evade traditional AI security checks that focus on datasets, model architectures, or training-time tampering, raising concerns about the robustness of deployed AI systems in sectors like transportation, healthcare, and finance. The work adds to a growing body of research on AI security showing that models are vulnerable not only to adversarial inputs and data poisoning but also to low-level hardware faults that can be weaponized. It underscores the need for more secure memory architectures, integrity checks on model parameters, and better deployment practices to protect AI models running on shared or untrusted hardware environments. "entities":["George Mason University","OneFlip","Rowhammer","Decrypt","deep learning models","self-driving cars","facial recognition systems"]}`

✨ AI-generated background, compiled from web sources — not editorial content.

More coverage

• Researchers found that "flipping" only one bit in memory is capable ... (ground.news)
• AI Can Be Hacked With a Simple 'Typo' in Its Memory, New Study ... (x.com)
• AI Can Be Hacked With a Simple 'Typo' in Its Memory, New Study ... (cryptonews.net)
• [PDF] Artificial Intelligence: An Introduction for Policymakers (mercatus.org)
• AI News at George Mason (gmu.edu)

Explore the topic

• University
• Claims
• Hacked
• AI