SShortSingh.
Back to feed

Shell Expansion Lets 'rm -rf' Commands Evade AI Guardrails, Study Finds

0
·33 views

A DEV Community article explains that simple text-matching safety filters can miss dangerous shell commands because the shell expands variables and quoting before execution. Its example, a variable set to root followed by a wildcard recursive delete, passes a regex-based check while expanding into a full system wipe. The author cites the GuardFall benchmark from June 30, 2026, which tested 11 autonomous coding agents and found most failed to block such bypasses. The piece argues that securing these agents requires parsing the shell's abstract syntax tree and resolving expansions statically before applying policy. It also describes context bloat from Model Context Protocol tool schemas, which inflates token costs and degrades model performance, and introduces an open-source tool called BlastRadius-Zero to address both problems.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingDEV Community ·

Unpatched VMware vCenter zero-day exploited by Chinese APT group

A critical, unpatched remote code execution flaw in VMware vCenter was actively exploited in a zero-day campaign attributed to a sophisticated Chinese advanced persistent threat group. Because vCenter serves as the central management plane for virtualised infrastructure, a compromise could expose every system it manages, including databases, domain controllers, and backup systems. The vulnerability requires no valid credentials and can run commands with root or SYSTEM privileges, allowing attackers to potentially gain full control of a virtual environment. The article argues that such a flaw demands emergency out-of-band response rather than standard patch cycles, along with pre-authorised mitigation paths and detection controls. Recommended steps include applying vendor patches immediately, isolating hosts and rotating credentials if no fix exists, and restricting management-plane access to trusted administrative sources.

0
ProgrammingDEV Community ·

Physical AI Requires Hardware and Software Co-Design, Says STATERA Library Chapter

Physical AI systems cannot be treated as models separate from their physical platforms, since perception, computation, decisions, and actuation form one integrated loop. Hardware factors such as timing, energy efficiency, reliability, and thermal limits directly shape what an AI system can achieve. System requirements should be derived from the physical task and operating environment, with models and computing platforms chosen together to avoid impractical or overly constrained designs. Sensors define what the AI can observe, while actuators define the physical actions it can take, making their characteristics central to architecture. The chapter emphasizes that increasing capability brings unavoidable tradeoffs among performance, power, cost, weight, and bandwidth.

0
ProgrammingDEV Community ·

Physical AI Co-Design: Why Robot Hardware and AI Must Be Engineered Together

The chapter argues that intelligent robots cannot be designed well by treating the physical platform, software, and AI model as separate parts, since their capabilities are tightly linked to the body and environment. It holds that system performance comes from integration rather than optimizing any single component alone. The first principle is that the mission, task, and operating environment should shape the architecture, with requirements for latency, robustness, power, and size derived from what the system must accomplish. The text also treats the robot's physical embodiment as a central design variable, since geometry, mass, actuators, and sensors determine what the AI can perceive and how it can act. Finally, it stresses selecting AI models and hardware together, because more capable models demand more compute, memory, and power, and hardware-aware model design is needed for real-world deployment.

0
ProgrammingDEV Community ·

Compute Platforms for Physical AI: Sizing Hardware Around Real-Time Robotic Workloads

This chapter from the STATERA Library's Physical AI hardware-software co-design volume argues that compute requirements should be derived from the full system rather than from hardware specifications alone. It holds that a robot must continuously convert sensor data into understanding, decisions, and control actions under real-world constraints, so workloads like perception, planning, and control must be sized together. Latency from sensing to action, including data movement and control, is presented as more important than peak theoretical throughput. The text also explains that heterogeneous sensor streams and different processor types, such as CPUs, GPUs, NPUs, and MCUs, shape the architecture needed for these demands.