AI agents

S&P Global CISO Dustin Wilcox and Docusign CISO Michael Adams argue that AI agents are a new class of non-human identity that existing identity models and inventory processes handle poorly. They advise CISOs to adopt an identity-first security model, starting with clean directories, enforced least privilege, reliable offboarding, and a complete inventory of non-human identities and their owners and authorizations.

Some people fear that frontier LLMs such as Claude Mythos and Anthropic's GPT-5.5 will lead to cybersecurity annihilation. Ari Herbert-Voss argues this could instead be an opportunity.

The article argues that agentic AI elevates the Tier 1 SOC analyst from performing repetitive investigation steps to overseeing a system of AI agents, shifting the role from operator to orchestrator. It states that a typical phishing alert investigation takes 20–30 minutes of pivoting across email logs, endpoint data and threat intelligence tools. The author contends that AI will expose the unmet workload in security operations rather than eliminate analyst jobs.

Choco, an AI-powered food and beverage distribution platform serving over 21,000 distributors and 100,000 buyers, built OrderAgent and VoiceAgent on OpenAI APIs to turn emails, texts, voicemails, images and phone orders into structured ERP orders. The company reports processing over 8.8 million orders annually and up to a 50% reduction in manual order entry.

At RSA Conference in March 2026, the industry focused on agentic AI, meaning AI systems that act as autonomous actors. The Cloud Security Alliance predicts a rise in simultaneous AI-powered attacks and urges defenders to fight AI with AI. The article argues that organizations should treat AI agents as identities and place them under identity threat detection rather than adding more point tools.

Orchid argues that enterprises face a delegation gap rather than a simple agent gap, since AI agents act on authority delegated by existing human and machine identities. The article says agents must be governed through the posture and intent of their delegators, with continuous observability of traditional identities as the foundation. The source text is cut off before it describes the full authority engine.

Israel-based Copperhelm emerged from stealth after raising $7 million in seed funding, led by TLV Partners, for its agentic cloud security platform. The platform deploys AI agents that monitor cloud environments, investigate threats, and execute remediation in real time, using a component called the Context Lake.

At Google Cloud Next '26, Google responded to Anthropic Mythos with a push toward agentic, AI-driven defense rather than another LLM. It announced three agents in Google Security Operations (threat hunting, detection engineering, and third-party context), expanded Wiz integration for AI development environments, and the Gemini Enterprise Agent Platform.

OpenAI released GPT-5.5, which it describes as its smartest model yet, with gains in agentic coding, computer use, knowledge work and early scientific research. It rolls out to Plus, Pro, Business and Enterprise users in ChatGPT and Codex, and GPT-5.5 Pro to Pro, Business and Enterprise users in ChatGPT, with API access to follow later.

Meta is installing a tool called Model Capability Initiative (MCI) on US-based employees' computers. It records mouse movements, clicks, keystrokes, and occasional screenshots in work-related apps and websites, and the data will train the company's AI models to interact with computers like humans, including automating work tasks. According to Reuters, the data won't be used for performance assessments.

Check Point and Google Cloud describe securing AI agents by controlling what agents are allowed to do, not only who can access systems. Google Cloud's Gemini Enterprise Agent Platform is presented as a centralized control point for agentic systems, covering identity, access, policy enforcement, and observability. The source text is truncated before any further detail.

OpenAI describes how it made agent loops on the Responses API 40% faster end-to-end, so Codex could run GPT-5.3-Codex-Spark at near 1,000 tokens per second instead of the roughly 65 TPS of prior flagship models. The gains came from caching rendered tokens and model configuration, removing intermediate network hops, speeding up safety classifiers, and adding a persistent WebSocket connection in place of repeated synchronous HTTP calls.

Deloitte's AI Institute 2026 State of AI report finds nearly 74% of companies plan to deploy agentic AI within two years, yet only 21% report a mature governance model for autonomous agents. The article argues that insecure agents can be manipulated into reaching sensitive systems and data, and that a centralized control plane governing agents, their permissions, policies, models and tools is needed to scale them safely.

Wiz has added Wiz Code plugins and skills, powered by the Wiz MCP server and WizCLI, that bring its security context into AI-native IDEs and coding agents. The tooling scans AI-generated code in real time and lets coding agents apply Green Agent remediation guidance, which can create pull requests. The announcement responds to frontier models such as Anthropic's Claude Mythos Preview, which the source says can autonomously discover and exploit zero-day vulnerabilities.