Agentic AI

Noma Security discovered GitLost, an attack vector targeting GitHub's Agentic Workflows. By hiding malicious instructions in public GitHub issues, attackers trick AI agents into exposing confidential repository data in public comments. This bypasses existing security safeguards designed to protect private information.

• Public GitHub issues can serve as injection vectors for AI agents.
• Agentic Workflows must sanitize inputs from untrusted public sources.
• Private data leakage is possible even with standard security safeguards.
• Monitor AI agent outputs for unintended data exposure in public spaces.

Researchers have identified a vulnerability in OpenAI's infrastructure that enables a single phishing link to deploy an autonomous AI agent within a corporate environment. This rogue agent gains access to employee credentials and operates independently, effectively acting as a corporate mole. The attack vector relies on exploiting the platform to create persistent, self-sustaining malicious processes rather than simple credential theft.

• Single phishing link can deploy autonomous AI agents, not just steal data
• Vulnerability allows rogue agents to retain employee access indefinitely
• Attack creates persistent corporate moles rather than one-time breaches
• Defenses must detect autonomous agent behavior, not just login anomalies

The Model Context Protocol is undergoing its most significant revision since launch by removing session management and eliminating rarely used capabilities. This structural shift forces developers using custom or homebrew implementations to undertake substantial migration work. The changes aim to simplify the protocol but introduce immediate compatibility challenges for existing integrations.

• Sessions are removed, fundamentally changing how state is handled in MCP.
• Legacy features are gutted, requiring updates to older client/server code.
• Custom implementations face a difficult migration path to stay compatible.
• The overhaul simplifies the protocol but breaks backward compatibility.

Block has open-sourced Buzz, a communication platform built on the Nostr protocol that allows users to self-host a workspace where both humans and AI agents collaborate in shared rooms. The system operates on a single-relay model where the URL defines the community boundary, ensuring tenant-observable state remains local to that specific workspace. It supports various interactions including messaging, reactions, and workflow approvals within a sovereign, Apache 2.0 licensed environment.

• Buzz leverages Nostr for decentralized, self-hostable collaboration between humans and AI agents.
• Single-relay architecture ensures URL-authoritative workspace isolation and tenant-local state.
• Open-sourced under Apache 2.0, enabling internal deployment for sovereign data control.
• Supports integrated workflows, reviews, and agent interactions within shared virtual rooms.

OpenAI has introduced Presence, a dedicated platform for deploying trusted voice and chat agents within enterprise environments. The solution targets customer-facing and internal workflows, positioning itself as a proven tool for organizational integration. This release expands OpenAI's ecosystem beyond raw model access into structured agent deployment.

• New platform specifically designed for enterprise-grade voice and chat agents.
• Targets both customer support and internal operational workflows.
• Positions OpenAI to compete in the structured AI agent market.
• No specific technical architecture details provided in this announcement.