AI Security News: November 13, 2026
Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled
1. Unit 42: One Operator Used the Open-Source Cairn Agent Framework With Zhipu and DeepSeek Models to Run 25 Intrusions Against 19 Organizations
Palo Alto Networks Unit 42 published an analysis on November 12 of Chinese-speaking attackers who used Cairn, an open-source AI penetration testing framework, for 25 operations against at least 19 organizations in August and September 2026, gaining unauthorized access or extracting data at 10 of them. Unit 42 tracks the activity as CL-STA-1238 and assesses with moderate confidence that it is China-nexus intelligence collection by one operator or a small team; targets included seven semiconductor and advanced-manufacturing firms, three defense or defense-policy bodies and four human rights and civil society groups. The attackers ran a customized fork from a Hong Kong server with an added exploitation skill pack, giving the model little more than a domain and a one-line goal; operations ran from 20 minutes to more than 50 hours. Thirteen operations used Zhipu GLM 5.2, eight DeepSeek V4 Flash and four DeepSeek V4 Pro, and neither vendor’s model refused after a Chinese-language preamble and a generated AUTHORIZATION.md file claimed the work was approved testing, according to Help Net Security’s summary of the report. The flaws exploited were ordinary ones (authentication bypasses, insecure direct object references, mass assignment, weak credentials, exposed cloud APIs, old CVEs), and one logged result was access to more than one million member records at a U.S. teachers’ union. Unit 42 concludes that autonomy raised the volume and coverage of exploitation but not its complexity. Source Help Net Security
2. Corelight: An MCP Server Pinned to One Release Ran With 15 Different Dependency Sets in Five Weeks
Help Net Security reported on November 13 on Corelight research showing that pinning an MCP server’s version does not pin what it installs. On an administrator workstation in Corelight’s lab, a server configured to a fixed release ran with 15 different sets of dependencies over five weeks with its configuration untouched: 23 of its 69 packages changed version, including the cryptography library, the JSON Web Token library and the MCP SDK. The cause was the uvx launcher installing from the package’s declared dependencies and skipping the project’s hash-pinned lock file, so about 60 transitive packages were resolved fresh from PyPI whenever the environment was rebuilt. The exposure is concrete because the servers on that workstation made more than 1,600 connections in a week to 11 management systems, including the firewall and the SIEM. Of 35 MCP servers Corelight reviewed from two vendor catalogs and its lab, 23 had a recommended install that pulls dependencies from a public registry on every launch, install or image build. Five servers frozen to a commit, installed once from a hash-pinned lock file and launched directly kept identical package lists across 457 starts in two weeks. Corelight’s own post is dated November 10. Source Corelight write-up
3. llm 0.37.2 Fixes Two Paths That Sent a User’s OpenAI API Key to Third-Party OpenAI-Compatible Endpoints
Simon Willison released llm 0.37.2 on November 11 to fix two medium-severity advisories reported by Daniel de Araujo, GHSA-v9h7-jj4g-f862 and GHSA-r639-rr7r-x72w; neither has a CVE ID or CVSS score listed. Both affect custom models defined in extra-openai-models.yaml with an api_base pointing at another provider. In the first, when the key named by api_key_name had never been stored, the tool fell back to OPENAI_API_KEY and sent it as the bearer token to the third-party endpoint; this affects versions 0.6 through 0.37.1. In the second, running a responses: true model with -o chat_completions 1 built a delegate chat model that kept the default openai key requirement and sent the stored or environment OpenAI key to api_base; this affects 0.32a2 through 0.37.1. In both cases nothing indicated to the user that the key had gone out, and the documentation said the OpenAI key would not be sent when api_base is set. Both are fixed in 0.37.2. Source GHSA-v9h7-jj4g-f862 GHSA-r639-rr7r-x72w
4. United Airlines Moves to Block Third-Party AI Agents From Booking and Managing Flights, Citing Shared Credentials
United Airlines said in a statement dated November 11, reported by TechCrunch on November 12, that it is restricting access by unauthorized third-party personal AI agents and assistants, both directly on its digital platforms and indirectly through travel agencies and aggregators. The airline’s stated reasons are that sharing account credentials with an assistant can expose sensitive information and lead to unintended purchases, account changes or use of MileagePlus miles, and that automated actions could violate FAA and DOT rules, for example the requirement that a human acknowledge at check-in that bags contain no banned substances. The work includes updates to booking, ticketing and distribution policies, and United plans to update its Digital Trust Center with guidance on the restrictions and on what to do when an agent takes an unintended action. The statement gives no technical detail on how agent traffic will be identified. TechCrunch notes that Bank of America urged customers the previous week not to share passwords with AI assistants. Source TechCrunch
5. Apify Says an Attacker Used a Compromised Credential to Obtain Account Credentials, and Revokes GitHub and Other OAuth Connections
Apify, the web scraping and automation platform whose Actors are often used as web-data tools for agents, disclosed a security incident on its status page, with updates posted on November 11 and 12. On November 9 and 10 an unauthorized third party used a compromised credential to obtain a number of account credentials and accessed the related Apify accounts; Apify detected the activity on November 11 at 12:13 UTC, closed the access path and reports no attacker activity since. As a precaution it rotated critical internal secrets and revoked some OAuth connections to external integrations, including GitHub, while stating that its investigation has not found those connections usable as an attack vector. Customers whose personal information or stored data was accessed have been notified by email, and users who build Actors from private GitHub repositories must reconnect GitHub once per account and organization. The number of affected accounts and how the initial credential was compromised have not been published; Apify has promised a postmortem after remediation. Source
6. PolicySniff Infers an Agent’s Confidential Approval Setting From GPU Timing on a Shared Accelerator
A paper submitted to arXiv on November 12 describes PolicySniff, a side channel that reveals which approval policy an LLM agent is running, such as whether it asks before a consequential tool call, without access to its prompts or responses. A co-resident attacker who knows the candidate settings and the task profiles both settings, then times only its own GPU kernels while the agent processes the task. The policy prefix is cached before capture, so the observed computation covers a task suffix that is identical under both settings. On an eager sparse mixture-of-experts runtime, where each selected expert runs as separate work, a frozen and blinded study identified the deployed setting with 94.17% balanced accuracy on its primary measure, and with 88.82% accuracy on tasks where both settings produce the same decision. The authors conclude that behavioral compliance does not imply configuration confidentiality and recommend enforcing authorization outside the model. The results are limited to the runtime and hardware setup studied. Source
7. MobileProbe Red-Teams Mobile GUI Agents by Placing Injections Along the Route Each Agent Takes
A paper submitted to arXiv on November 12 introduces MobileProbe, a red-teaming framework for mobile GUI agents that operate phone apps from screen content. The authors argue that existing benchmarks understate risk because they pin a hand-written payload to a fixed step of a fixed task, so an agent that takes a different route and never reaches the injected page is recorded as safe without being tested. MobileProbe finds attacker-controllable content surfaces through automated pentesting, locates vulnerable decision points along the agent’s own trajectory, and refines context-aware payloads in an evolutionary loop filtered by a surrogate model. The accompanying benchmark has 83 long-horizon attack instances across 12 mobile applications, each labeled with a confidentiality, integrity or availability goal. On three agent architectures it reports higher attack success rates than human-written payloads, zero-shot LLM payloads and earlier automated red-teaming baselines; the abstract does not give the absolute rates. Source