🧭 Claude Partner Network Goes Tiered: Services Track and Partner Hub Now Live

Anthropic officially launched the Services Track and Claude Partner Hub — two additions to the Claude Partner Network that formalise how consulting firms and SIs prove and advertise their Claude expertise to enterprise buyers. The partner network itself launched in March with a $100 million investment in training and shared marketing; more than 40,000 firms applied and over 10,000 consultants have since earned a Claude certification. The new tier structure and portal surface that credibility in a way prospective enterprise customers can act on directly.

Services Track: three tiers, earned not bought

Unlike many vendor partner programs where tiers are purchased via fees, the Claude Services Track tiers reflect actual delivery work:

• Select — 10 certified Claude practitioners, 2 production client deployments, 1 public customer endorsement.
• Preferred — 100 certified practitioners, 15 active customer deployments, 3 published endorsements.
• Global Premier — 1,000 certified practitioners, deployments with 100+ customers across 3+ regions, 15 public endorsements, and a jointly developed business plan with named executive sponsors from Anthropic.

Requirements are published and standing is refreshed daily in the Partner Hub dashboard, so there is no ambiguity about where a firm sits or what it needs to advance.

Claude Partner Hub: customer-facing directory + MCP connector

The Partner Hub serves two audiences. For partner firms, it's a live dashboard showing certification counts, deal registrations, and gap-to-next-tier metrics. For enterprise buyers, it's a searchable directory showing which firms have real deployments in a given industry or use-case. Both audiences can query the Hub through a new MCP connector — for example, a firm's internal Claude instance can ask "where does our practice stand against Preferred tier?" and get a data-backed answer directly from the Hub.

🧭 Anthropic Publishes Dual Cyber Reports: AI Threat Mapping and the First AI-Orchestrated Espionage Campaign

Anthropic released two significant security research publications simultaneously. The first, What we learned mapping a year's worth of AI-enabled cyber threats, examines 832 accounts banned for malicious activity between March 2025 and March 2026, mapping them onto the MITRE ATT&CK framework. The second, Disrupting the first reported AI-orchestrated cyber espionage campaign, describes a state-sponsored operation detected in September 2025 where AI autonomously conducted 80–90% of attack operations against roughly 30 global targets.

Key findings from the year-long threat analysis

• The share of medium- or high-risk actors using AI for cyber operations rose from 33% to 56% — a 1.7× increase year-over-year.
• 67.3% of the 832 banned accounts used AI primarily for malware writing — the most prevalent activity.
• AI is being deployed at the later, more complex stages of attack chains (exploit development, evasion, lateral movement), not just reconnaissance — a shift in how defenders must model AI-assisted threats.
• Anthropic found that MITRE ATT&CK lacks adequate techniques to describe AI-enabled attack behaviour, and is in discussion with MITRE about updating the framework accordingly.

The espionage campaign: GTG-1002

In September 2025, Anthropic detected a sophisticated campaign attributed to a Chinese state-sponsored group designated GTG-1002. The operation targeted roughly 30 entities including large tech companies, financial institutions, chemical manufacturers, and government agencies. Against one targeted technology company, the threat actor directed Claude Code to independently query databases, extract data, parse results for proprietary information, and categorise findings by intelligence value — autonomously, at physically impossible request rates that made the non-human orchestration evident. Anthropic banned accounts as they were identified, notified affected entities, and coordinated with authorities.

🧭 Claude Code Hardens Its Defaults: Streaming Tools Always On, MCP Session IDs & Sharper Exfiltration Detection

The latest Claude Code release ships several security and reliability improvements. The changes are incremental but operationally significant — particularly for teams running Claude Code in automated pipelines or connecting it to internal MCP servers.

What changed and why it matters

• Streaming tool execution is now always enabled — previously it was behind a feature flag and disabled when telemetry was off or when running on Bedrock, Vertex, or Foundry. Tools now stream consistently across all deployment paths, closing a subtle reliability gap that could produce different behaviour in local vs. CI vs. cloud environments.
• MCP session IDs injected into subprocess environment — stdio MCP server subprocesses now automatically receive CLAUDE_CODE_SESSION_ID and CLAUDECODE=1 in their environment. MCP servers can use these to correlate logs across a session, build session-scoped caches, or route tool calls to tenant-specific resources without any additional integration work.
• Improved data-exfiltration detection in auto-mode — the classifier now more aggressively identifies bulk transfers of repository contents, tightened specifically in response to patterns identified in the GTG-1002 espionage investigation. Any task that attempts to read and transmit large portions of a codebase outside the working directory is now more likely to be flagged and paused for human review.
• MCP server approval UX fix — claude mcp list/get now shows servers defined in unapproved .mcp.json files as Pending approval instead of auto-approving them when output is piped. This prevents unreviewed MCP servers from silently gaining access in non-interactive contexts such as CI pipelines.

Practical steps for teams using Claude Code in CI

# Before this release, streaming was only enabled when telemetry was on.
# Now it's unconditional. If you had workarounds like:
#   CLAUDE_CODE_DISABLE_STREAMING=0 claude ...
# those env-var overrides are no longer necessary and can be removed.

# To take advantage of session IDs in a custom MCP server:
# In your MCP server startup handler:
import os
session_id = os.environ.get("CLAUDE_CODE_SESSION_ID", "unknown")
# Use session_id to namespace caches, logs, or tenant state

# Audit your .mcp.json files — any that weren't explicitly approved
# will now appear as "Pending" in `claude mcp list`. Review and
# approve them explicitly before the next CI run:
claude mcp approve <server-name>