Safety & Security · Thu 12 Nov

AIAI-generated summary CSO Online reports that coding agents running on developer laptops connect to MCP servers and use credentials to access repositories and services. CSO Online notes that security professionals face challenges governing autonomous software that holds active pathways into production.

Code Agents MCP Model Security

AIAI-generated summary Alibaba Cloud Community reports that the Higress project has integrated Qwen3Guard-Gen as a Wasm plugin to handle generative AI safety moderation at the gateway level. The default plugin configuration calls Qwen/Qwen3Guard-Gen-4B and maps official result levels to two gateway policy settings.

Alibaba Qwen Small Models Guardrails

AIAI-generated summary Chainstack revealed an AI-powered security breach that exposed customer names, emails, and password hashes after an attacker used the ARTEX AI agent framework. The company stated that the incident has been contained and is currently asking customers to rotate their node credentials and change their passwords.

Model Security

AIAI-generated summary Modern generative artificial intelligence and large language models secretly create and store psychological impressions of users during ordinary conversations. These hidden profiles form by default without user knowledge or consent and can quietly influence subsequent interactions.

Data Privacy Guardrails Chatbots & Assistants

AIAI-generated summary Anthropic launched Project Glasswing in April using the Claude Mythos Preview engine to find software vulnerabilities. Patrick Garrity reports that the project found only two exploited vulnerabilities and produced many low-risk findings.

Anthropic Code Review Hallucinations

AIAI-generated summary AI companies have stated that their technology poses existential risks, including statements warning that it could end humanity in the near future. Anthropic released a report detailing real-life incidents of AI misuse, leading to genuine concern among some experts.

Anthropic

AIAI-generated summary Help Net Security reports that OWASP ranks improper offboarding as the top risk for non-human identities like AI agents. Help Net Security notes that Sprinto customers bring AI systems into their compliance program by placing each one in a registry with a named owner.

Guardrails

AIAI-generated summary Cequence reports that automated credential stuffing operations used the identity of a consumer AI agent on an online marketplace during a week in October. Cequence advises organizations to monitor session behavior rather than relying entirely on conventional bot scores to detect malicious activity.

Guardrails

AIAI-generated summary Apple has shared a detailed document explaining Secure Exclaves, which is the security and privacy toolset behind features like Siri Recaps. According to the report, these components allow video and audio data to be analyzed without system access or saving.

Apple Data Privacy Model Security