← Back to Blog

Agentic AI this week (Apr 24 - May 1, 2026)

Weekly roundup: GPT-5.5 ships with 1M-token context and native agentic execution, RSAC 2026 makes agentic security the defining theme with 1.3B agents projected by 2028, Google unveils agentic SOC with Wiz integration, Fortinet expands FortiAI with MCP support across unified FortiSOC, Check Point integrates AI Defense Plane with Google Cloud, Cloudflare ships Project Think with durable actor runtime, MCP crosses 150M installs with RSAC vendor adoption surge

📅 May 1, 2026 ✍️ Authored by AI ⏱️ 16 min read 📁 AI

AI Agentic MCP

Agentic AI this week (Apr 24 - May 1, 2026): GPT-5.5, RSAC 2026, and the Security Vendor Land Grab

This was the week everything converged. OpenAI shipped GPT-5.5 with native agentic execution and a 1M-token context window. RSAC 2026 made agentic AI security the defining conference theme, with the industry projecting 1.3 billion AI agents by 2028. Google unveiled an agentic SOC stack built on its $32B Wiz acquisition. Fortinet expanded FortiAI with MCP support across the unified FortiSOC. Check Point integrated its AI Defense Plane with Google Cloud. Cloudflare shipped Project Think, a durable actor-based runtime for production agents. And MCP crossed 150 million installs with three major security vendors announcing protocol support at RSAC. The agentic era is no longer arriving. It arrived.

GPT-5.5 launches with native agentic capabilities

OpenAI released GPT-5.5 on April 23, with API access (GPT-5.5 and GPT-5.5 Pro) following on April 24. The model ships with a 1 million token context window across ChatGPT Plus, Pro, Business, and Enterprise tiers, and processes text, images, audio, and video in a single unified architecture, a first for OpenAI.

Benchmark scores show the agentic focus: Terminal-Bench 2.0 at 82.7% and FrontierMath at 51.7% (levels 1-3) and 35.4% (level 4). API pricing holds at $5/$30 per million input/output tokens for the standard tier, with the Pro tier at $30/$180. Key agentic improvements include better error recovery mid-task, more efficient tool calls, and coherence over longer contexts.

Why it matters: GPT-5.5 is explicitly positioned as an agentic model. It can write code, research online, analyze data, create documents, operate software, and move across tools until a task is finished. The 1M-token context window and unified multimodal architecture represent OpenAI's architectural bet on long-running autonomous agent workflows. The competitive landscape at the frontier is now a three-way race between GPT-5.5, Claude Opus 4.7 (87.6% SWE-bench), and Gemini, with each model optimizing for different agentic strengths. OpenAI is betting on context length and multimodal breadth. Anthropic is betting on coding reliability and tool-use consistency. The market will decide which tradeoffs matter more in production.

RSAC 2026 makes agentic AI security the defining theme

RSA Conference 2026 (April 28 through May 1, San Francisco) made agentic AI security the dominant theme across every keynote and expo floor. The industry projection that anchored the conversation: 1.3 billion AI agents by 2028. Three security pillars emerged as consensus requirements: agent discovery, runtime protection, and identity governance for non-human actors.

Mandiant's M-Trends 2026 report delivered the stat that reframed urgency: adversary hand-off times have collapsed to 22 seconds from initial access to second-stage deployment. CrowdStrike, Google/Wiz, Check Point, Fortinet, Cisco, and Palo Alto Networks all shipped agentic SOC tools at the conference. The key threat vectors highlighted across vendor presentations were prompt injection, model poisoning, shadow agents, and unauthorized model access.

Why it matters: RSAC 2026 marked the moment agentic AI security went from buzzword to architectural imperative. Every major vendor now treats AI agents as a primary attack surface requiring discovery, governance, and runtime protection. The 22-second adversary hand-off time is particularly alarming in the context of autonomous agents: if human defenders already cannot keep up with human attackers at 22 seconds, they certainly cannot govern autonomous agents that operate at machine speed without automated security controls. The coordinated vendor response suggests the industry has internalized this reality.

Google unveils agentic SOC with Wiz integration and Gemini Agent Platform

Google Cloud used RSAC to launch the most comprehensive agentic security stack in the industry, combining its Gemini AI, the completed $32B Wiz acquisition, and Mandiant threat intelligence into a unified offering.

The Gemini Enterprise Agent Platform shipped with three new enterprise capabilities: Agent Identity (access management for non-human agents), Agent Gateway (policy enforcement for MCP and A2A traffic), and Agent Registry (centralized discovery and inventory). Wiz introduced red, blue, and green security agents for multi-cloud AI application protection. Google Security Operations now has Triage and Investigation agents in preview that autonomously investigate alerts, gather evidence, and produce verdicts. MCP server support in Security Operations went GA. Model Armor was extended to cover agentic risks including prompt injection and tool poisoning. A Dark Web Intelligence agent analyzes millions of daily external events with 98% accuracy using Gemini models.

Why it matters: Google is assembling the most comprehensive agentic security stack by combining its own AI, the $32B Wiz acquisition, and Mandiant threat intelligence. The Agent Gateway natively understanding MCP and A2A protocol traffic positions Google as the infrastructure layer for enterprise agent deployments. The fact that Google is building security both for agents (protecting agentic workloads via Wiz and Model Armor) and with agents (automating SOC workflows via Gemini) gives it coverage across both sides of the market.

Fortinet expands FortiAI with MCP support across unified FortiSOC

Fortinet expanded FortiAI across FortiAnalyzer, FortiSIEM, FortiSOAR, and FortiSOC at Accelerate 2026 and RSAC 2026, moving beyond interactive copilots to fully agentic execution. A dedicated FortiAI agent now automates alert triage, investigation workflows, and threat hunting across the unified platform.

The most significant technical announcement is explicit MCP support that maintains shared context and execution continuity across detection-investigation-response workflows. FortiSOC, the cloud-delivered unified SOC platform that merges FortiAnalyzer, FortiSIEM, FortiSOAR, and FortiTIP into a single service, was previewed with agentic workflows at its core. New FortiEndpoint capabilities and managed services were also announced.

Separately, FortiGuard Labs released the 2026 Global Threat Landscape Report on April 30, documenting a 389% increase in confirmed ransomware victims (reaching 7,831 globally), fueled by crime-as-a-service kits and agentic AI tools. The report describes cybercrime now operating as a system with "shadow agents" compressing the entire attack lifecycle, and a 79% increase in comprehensive data set theft enabled by agentic AI.

Why it matters: Fortinet's MCP support means FortiAI agents can maintain shared context across the entire detection-investigation-response chain, carrying investigation state as the agent moves between tools. This is architecturally different from the traditional SOAR approach of scripted playbooks. The FortiSOC consolidation into a single cloud platform with agentic workflows represents Fortinet's strategic shift from tool-per-function to unified agentic security operations. The 389% ransomware increase documented in the threat report provides the urgency narrative: manual SOC workflows cannot scale against this threat velocity.

Check Point integrates AI Defense Plane with Google Cloud

Check Point announced a deepened integration between its AI Defense Plane and Google Cloud's Gemini Enterprise Agent Platform on April 22, extending the platform built on the Lakera and Cyata acquisitions to protect agentic workloads running on Google Cloud infrastructure.

The integration provides five capabilities: discovery, governance, observability, runtime control, and continuous validation across the AI execution lifecycle. AI Application and Agent Security discovers where AI is present, what data and tools it accesses, evaluates behavior, and governs permissions and trust relationships. AI Red Teaming capabilities enable continuous adversarial testing of prompts, reasoning paths, workflows, tool use, and agent behavior. The platform integrates with Google Cloud's Agent Gateway and Agent Registry for centralized whitelist/blacklist policies and real-time prompt inspection. Availability is targeted for late June 2026.

Why it matters: Check Point's framing of "outcome control" replacing "access control" captures the fundamental shift in agentic security. When AI agents act autonomously, you need to govern what they do, not just where they connect. The Google Cloud partnership positions Check Point as the security layer for Google's agent ecosystem, mirroring the pattern where security vendors align with specific cloud platforms. Palo Alto Networks countered at RSAC by announcing the Portkey acquisition to build its own AI Gateway into Prisma AIRS. The security vendor land grab for the agentic AI market is now in full swing, with each major vendor staking out different architectural positions.

Cloudflare ships Project Think: durable actor runtime for production agents

Cloudflare shipped Project Think during Agents Week (April 13-17), introducing a durable actor-based runtime for AI agents that fundamentally changes how production agents manage state and execution. The platform includes Fibers for crash recovery and checkpointing, sub-agents with isolated SQLite databases, persistent sessions with full-text search, and self-authored extensions that allow agents to write their own tools at runtime.

Additional Agents Week launches include Cloudflare Mesh, the first private networking solution built specifically for AI agents (integrating with Zero Trust and Workers VPC), Dynamic Workers with isolate-based sandboxed runtime starting in milliseconds (100x faster than containers), Artifacts (Git-compatible storage for agent state), and Sandboxes GA (persistent Linux environments). In total, Cloudflare shipped 20+ new features in a single week.

Why it matters: Project Think's shift from stateless orchestration to durable actor-based execution addresses the hardest infrastructure problems in deploying agents at scale. Crash recovery means an agent that fails mid-task can resume from a checkpoint rather than starting over. Sub-agents with isolated databases enable compositional agent architectures without shared state corruption. Self-authored extensions, where agents write their own tools at runtime, is a capability that blurs the line between agent and platform. The 100x speed advantage over containers for Dynamic Workers positions Cloudflare as the edge-native alternative to centralized cloud agent runtimes.

A2A protocol solidifies with Microsoft Copilot Studio going GA

Microsoft Copilot Studio made multi-agent orchestration generally available in April 2026 with native A2A protocol support, Fabric integration, and autonomous agentic actions in Word, Excel, and PowerPoint. Copilot Studio now supports cross-SDK orchestration and third-party agent delegation via A2A.

The A2A protocol (hosted by the Linux Foundation) has grown to 150+ supporting organizations with its GitHub repository surpassing 22,000 stars. The SDK ecosystem expanded from one Python implementation to five production-ready languages: Python, JavaScript, Java, Go, and .NET. A2A v1.0 is now in production at Microsoft, AWS, Salesforce, SAP, and ServiceNow.

Why it matters: Microsoft making A2A-based multi-agent orchestration GA in Copilot Studio, with autonomous actions inside Office applications, signals that multi-agent systems are moving from experimental to enterprise production. An agent that can autonomously update Excel spreadsheets, modify Word documents, and adjust PowerPoint presentations based on cross-agent communication is a fundamentally different product from a chatbot. The five-language SDK ecosystem removes the adoption barrier, and the Linux Foundation stewardship provides the governance stability that enterprise buyers require.

MCP ecosystem surges with RSAC vendor adoption and Dev Summit

The Model Context Protocol ecosystem hit an inflection point this week. Three major security vendors, Google, Fortinet, and Check Point, all announced MCP integration at RSAC 2026, confirming MCP as the de facto standard for agent-to-tool communication in enterprise security. Google's Agent Gateway natively understands MCP protocol traffic. Fortinet's FortiAI uses MCP for shared context across SOC workflows. Check Point's AI Defense Plane governs which MCP servers agents can access.

The Agentic AI Foundation (AAIF) held the MCP Dev Summit North America in New York City with approximately 1,200 attendees. Python SDK and Java SDK both received updates on April 30. MCP has now crossed 150 million cumulative installs. Governance expanded with new Lead Maintainer Den Delimarsky (Anthropic) and Core Maintainer Clare Liguori contributing to the Triggers and Events working group.

Why it matters: MCP is maturing from protocol specification into critical production infrastructure. The 1,200-attendee dev summit, the 150M install milestone, and the fact that three major security vendors independently chose to integrate MCP at the same conference confirms that MCP is no longer just Anthropic's protocol. It is becoming the assumed connectivity layer for the entire agentic AI ecosystem. The security vendor adoption is particularly significant because it means MCP traffic will now be inspected, governed, and monitored by enterprise security stacks, which accelerates enterprise MCP adoption by addressing the security concerns that have slowed deployment.