Ox.Security Critical Flaw in DeepSeek Harness Allows AI Agents to Escape Sandbox
Article Content
- •CVE-2026-82533 allows AI agents to escape their sandbox.
- •The vulnerability requires no authentication and can be exploited with a single command.
- •DeepSeek Harness versions 0.1.1-rc.2 and earlier are affected.
A critical vulnerability (CVE-2026-82533) was discovered in DeepSeek Harness, an open-source tool for AI coding agents, enabling a sandboxed agent to disable its own confinement using a single shell command. The flaw stems from the tool's unauthenticated API, which relied on the client-supplied 'Host' header for trust verification, allowing an agent to elevate its session to 'danger-full-access' without approval prompts. This vulnerability affects default installations of DeepSeek Harness, which could expose sensitive data if the API was accessible over a network. The issue was reported by OX Research and remediated in version 0.1.2-alpha.1, released on August 27, 2026. The flaw has been rated 9.4 out of 10 by VulnCheck, indicating its severity. Users are urged to update to the fixed version to mitigate risks.
Ask AI about this cluster
Answers cite the sources they use
Timeline
More articles in this cluster (3)
Following this threat?
Track OX Research and CVE-2026-82533 in your own feed — you're alerted when they show up in new reporting, leak sites or exploitation.
Free account · no card needed
Continue Reading
Critical BeyondTrust 9.3 Vulnerability Actively Exploited by Threat Actor A critical vulnerability (CVE-2026-102764) in BeyondTrust Remote Support has been discovered and is actively being exploited by the threat group ARMORY CASTLE, reportedly linked to China. The vulnerability, which has a CVSS score of 9.3, allows unauthenticated attackers to upload files and potentially compromise host… 
Lazarus APT Launches Operation FlashHole Using AI and Post-Quantum Cryptography Kaspersky has identified a cyberespionage campaign, dubbed Operation FlashHole, attributed to the Lazarus APT group, active since July 2025. The campaign has targeted at least three organizations in South Korea's construction, software, and wire-communication sectors. Attack methods include spear-phishing emails…