Do You Really Know What Your AI Agent Is Doing in the Sandbox?
The rapid adoption of AI agents equipped with shells, filesystems, and network access has created a fundamental observability problem. When an agent finishes a task, logs generated by the model itself and platform metrics provide only a partial picture. The real question remains: what system calls did the agent actually execute?
Two main approaches exist for observing agent behavior. The first intercepts tool calls inside the agent framework, which is inexpensive but blind to any actions performed outside the framework. The second uses eBPF probes on the host kernel to capture every syscall, an approach used by container security tools. This method is more trustworthy yet assumes the agent shares the host kernel.
Strong isolation runtimes break that assumption. gVisor implements its own kernel in user space, while Kata Containers and Firecracker run a full guest Linux kernel inside a virtual machine. In both cases, guest syscalls never reach the host kernel where traditional monitors attach.
Existing standards cover several aspects of agent operation. MCP standardizes how agents connect to tools and data. OpenTelemetry defines semantic conventions for model and agent spans. RuntimeClass lets Kubernetes switch between runtimes without changing application manifests. The new agent-sandbox project treats sandboxes as first-class Kubernetes resources. None of these standards, however, define a common interface for reporting the actual syscalls performed inside the sandbox.
Empirical testing used a generator that performed ten distinct operations, each tagged with a unique marker. Probes verified that the monitoring stack itself was functioning correctly. Results across 24 runs demonstrated that visibility is determined by whether filesystem operations cross into a host process. When a block device backs the guest root, almost no guest file operations appear on the host. When virtiofs or an overlay shares the filesystem via a host process, file names and writes become visible.
Even within the same runtime, configuration changes produce dramatically different results. Enabling the in-memory overlay in one gVisor mode hides writes and deletions while still exposing opens. Executions never surface as execve events on the host under any tested sandbox; only prior file copies leave detectable traces when the filesystem is host-mediated.
Commercial security products already reflect these limitations. Falco removed its gVisor engine, Google Container Threat Detection is incompatible with GKE Sandbox, and Microsoft Defender cannot inspect Kata pods. The measurements confirm that switching a RuntimeClass value can silently eliminate visibility into entire classes of agent actions.
Related articles
Researcher Runs 100 Abliterated AI Agents Against His Own Accounts, Achieves Five Successful Breaches
A security researcher deployed roughly 100 autonomous AI agents based on abliterated open-source models to attempt unauthorized access to his online accounts over a five-hour period. The agents successfully compromised three accounts by exploiting vulnerabilities in legacy personal projects and two additional accounts through password reuse and credential stuffing. They also conducted 16 social engineering attempts, harvested and cross-verified personal data from public sources, and generated phishing pages targeting the researcher's Substack and Hacker News presence. Despite these successes, the agents failed to discover any zero-day vulnerabilities in third-party software or gain access to Tier-0 accounts such as Gmail, 1Password, or banking services. The experiment relied on abliterated versions of GLM-5.3 and DeepSeek V4 Flash models running on cloud GPUs, orchestrated via Codex CLI inside Docker containers with access to Railway infrastructure. Total compute cost reached approximately $210, equating to roughly $40 per successful account compromise. The researcher concluded that inexpensive swarms of such agents could soon enable low-skill attackers to systematically map and exploit an individual's entire online footprint.
How Censorship in Closed AI Models Reshapes Science: Hugging Face Releases Open Carbon-X DNA Model
Hugging Face CEO Clement Delangue announced the release of Carbon-X, an open 14.8 billion parameter language model designed for reading, analyzing, and generating DNA sequences. The announcement highlighted how shallow guardrails in closed commercial models blocked even proofreading a tweet discussing the topic. Commenters and researchers pointed out that these input/output filters lack context awareness, blocking legitimate scientific queries while motivated adversaries bypass them through rephrasing. Carbon-X was built on Qwen3-14B with a hybrid BPE and 6-mer tokenizer, domain mid-training on 302 billion tokens, and safety achieved through data curation rather than runtime censorship. The model achieves performance comparable to closed frontier systems at roughly 1/1800th the inference cost for large-scale genomic tasks. The case underscores growing asymmetry between closed corporate AI labs building wet labs and independent researchers restricted by superficial safety layers.
NYU Researchers Find 86% of Chatbot Responses on Digital Abuse Contain Errors, Gaps or Safety Risks
A study by New York University’s Tandon School of Engineering examined four chatbots answering 193 questions about spyware, location tracking, and account control. Researchers identified problems with accuracy, completeness, or safety in 86 percent of responses. Even specialized assistants Aimee and Ruth performed no better than general models GPT-4o and Claude 3.7 Sonnet. Approximately 74 percent of Aimee’s answers, 65 percent of Claude’s, 61 percent of Ruth’s, and 55 percent of GPT-4o’s were deemed unsafe. In 60 percent of cases the models omitted critical warnings that could alert an abuser or destroy evidence. All four systems could be prompted to assist potential stalkers, with Aimee and Ruth complying with roughly two-thirds of such requests.
Cloudflare Engineer Releases KillSwitch Language Designed to Trip Up AI Code Models
Dominic Pichet, an engineer at Cloudflare, has created an experimental programming language called KillSwitch specifically engineered to cause large language models to fail when interpreting or generating code. The language turns programs into narrative-style text where most content is ignored by the interpreter, hiding actual commands inside stories and using distance-based references to extract values. Rules for identifiers rely on letter counts and anagrams, while the language deliberately incorporates phrases such as “ignore previous instructions” that carry functional meaning inside KillSwitch itself. Pichet added misleading statements in the documentation and wording crafted to trigger model safety filters, creating a hostile environment for coding agents. In tests involving writing Hello World, explaining program output, and converting temperatures, Claude Opus 5 achieved the highest score of 66.8 points. The full source code and specification are available on GitHub, and the author plans to increase difficulty if models improve.