Loading vulnerability…
CVE-2026-82533 is an authentication-bypass vulnerability in DeepSeek Harness versions before 0.1.2-alpha.1. The local HTTP agent-control API makes its trust decision from the client-supplied HTTP Host header, accepting loopback or configured trusted-host values, instead of verifying the actual peer address of the TCP connection. An untrusted client can spoof an accepted Host value to access privileged control-plane functionality. In the default configuration, a sandboxed agent can access the loopback API and modify its own session policy to danger-full-access, disabling sandbox restrictions and approval prompts. The API can also expose stored conversation data to unauthorized callers.
Mallory correlates every CVE against your assets, your vendors, and active adversary campaigns. Know which vulnerabilities matter for you, not just which ones are loud.
What it means. What to do now. Patch path, mitigations, and the assume-compromise checklist.
What an attacker gets, and what they’ve been doing with it.
If you can’t patch tonight, do this now.
Patch, then assume compromise.
No public exploits tracked yet. Mallory keeps watching.
No public exploit code observed for this vulnerability.
14 sources tracked across advisories, community write-ups, and news. New activity surfaces here as Mallory finds it.
A critical sandbox-escape and local web-interface authentication flaw in DeepSeek Harness. An agent induced by attacker-controlled text could invoke the locally reachable interface, set its own session to danger-full-access, and execute commands outside its workspace without an approval prompt. The affected interface trusted a client-controlled Host header rather than verifying connection origin and had no authentication; it also exposed session-log retrieval to reachable callers.
Critical CVSS 9.4 authentication/trust-validation flaw in DeepSeek Harness's local agent-control API. Trust was determined from a spoofable client-supplied Host header rather than the actual peer address. Because the default sandbox permitted loopback networking and routine shell commands, an agent induced to execute attacker-controlled text could disable its own confinement without credentials, network exposure, or approval. Exposed instances could also be controlled remotely without authentication.
An authentication-bypass vulnerability in DeepSeek Harness's local HTTP control-plane API. The server trusts a client-supplied Host header rather than validating the actual TCP connection origin, enabling unauthenticated attackers to control agents, execute privileged danger-full-access commands, weaken approval policies to unconfined execution, and access stored conversations.
Query your assets running an affected version, and investigate the blast radius.
Every observed campaign linking this CVE to a named adversary.
Malware families riding this exploit, with evidence and IOCs.
YARA, Sigma, Snort, and vendor rules, auto-deployed to your SIEM.
Cross-references every affected SKU, including bundled OEM variants.
Community discussion across Reddit, Mastodon, and other social sources.