Do Sandbox Restrictions Actually Work for AI Agents Running in Linux and gVisor?
An experienced security researcher has published a detailed case study asking whether access-control mechanisms inside Linux sandboxes used by AI agents actually deliver the expected restrictions. The work focuses on scenarios where an agent runs inside an isolated environment—built on container runtimes or virtual machines—yet individual tools launched by the agent must operate with narrower rights than the full session.
The author explains that simply instructing an LLM “do not modify files” does not reduce process privileges. Instead, a trusted wrapper must apply concrete Linux controls such as Landlock, classic BPF socket filters, or CGROUP_DEVICE programs before handing control to the tool. Because these controls may behave differently under gVisor’s Sentry kernel, each mechanism must be validated from inside the sandbox with the exact privileges the wrapper will possess.
Experiments were performed on Ubuntu 22.04 (Linux 6.8.0-138-generic) and two gVisor releases using both the runsc do command and full OCI bundles. With elevated capabilities, Landlock ABI version queries returned ENOSYS in both gVisor releases, while the host kernel reported ABI 4. Kernel selftests confirmed the absence of Landlock handlers. When a CGROUP_DEVICE program was loaded and attached to a cgroup v2 hierarchy, gVisor accepted the operations and reported the program as attached, yet /dev/null and /dev/zero remained accessible—contrary to the host, where EPERM was correctly returned.
A classic BPF filter attached via setsockopt(SO_ATTACH_FILTER) also succeeded without privileges. On the host the filter dropped UDP datagrams as intended; under gVisor the same packets continued to arrive. The discrepancy persisted across network modes (--network=none and --network=host) and user IDs.
The researcher concludes that three distinct failure modes must be distinguished: absence of the mechanism, insufficient privileges to configure it, and successful configuration that nevertheless produces no effect. A practical acceptance procedure is proposed: verify that the target action succeeds without the policy, confirm the policy is accepted, then confirm the action is blocked afterward. The same sequence must be executed with the exact rights the AI tool will receive, and the entire test repeated after any change to kernel, runtime, image, or mount configuration.
If a mandatory restriction cannot be shown to work, the tool must not be launched. The article provides the full test harness and instructions in a public repository so practitioners can reproduce the checks for their own AI-agent sandboxes.
Related articles
TenetBench Benchmark Tests LLMs for Compliance with Professional Ethical Codes
Researchers from the Institute of Information Security Issues at ISP RAS and the Institute of Cybersecurity and Information Protection at RANEPA have developed TenetBench, a benchmark designed to evaluate how well large language models adhere to specific norms in professional codes of ethics. The benchmark generates realistic work-related prompts that may conflict with rules from documents such as the Code of Professional Ethics for Advocates or journalistic standards on undisclosed advertising. Each test case pairs a conflicting request with a neutral counterpart to measure whether models refuse, warn, or still provide prohibited content. Evaluation separates task completion, norm violation detection, and presence of explicit refusals using an LLM-as-a-Judge approach validated against human annotators. The full dataset contains 3,433 prompt pairs derived from 15 professional codes, with a public subset of 2,235 pairs from 10 documents. Human validation confirmed 99.5% of generated examples as fully plausible work scenarios.
Solar to Showcase Solar AI Firewall for Business Protection Against AI Risks at SOC Forum 2026
GK Solar will present its new Solar AI Firewall solution at SOC Forum 2026 to help organizations control interactions with AI services and prevent confidential data leaks through employee prompts. The product addresses growing risks from staff sending corporate data to neural networks and AI agents gaining access to internal systems. The forum takes place on 27-28 October in Moscow's Lomonosov cluster, with the official launch of Solar AI Firewall scheduled during the event. Solar will also highlight practical risks of AI adoption, including data transmission to external services, use of public models by employees, and new attack surfaces in corporate infrastructure. The solution will be demonstrated as part of an integrated ecosystem for access, network, and data protection. In addition, the company plans to introduce JSOC ONE, a new line of Solar JSOC services tailored for mid-sized businesses that require 24/7 monitoring without building their own SOC.
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.