Key Signals

• GitHub tightens Copilot individual plans: signups paused, usage limits reduced, Opus models restricted. New signups for Copilot Pro, Pro+, and Student plans are paused to prioritize service quality for existing customers, while tighter usage limits are imposed across individual tiers. Opus models have been removed from Pro entirely; Opus 4.7 remains on Pro+ only, with Opus 4.5 and 4.6 also scheduled for removal. This signals a shift toward sustainability-driven pricing and positions Pro+ as the premium individual tier, while Copilot Free remains open for new users. [1]

• Amazon invests $5B more in Anthropic, securing a $100B AWS cloud commitment over 10 years. Amazon's total investment in Anthropic now reaches $13 billion, with the deal structured around Trainium2 through Trainium4 custom AI accelerator chips and up to 5 GW of compute capacity. This mirrors Amazon's earlier $50B contribution to OpenAI's $110B round, establishing a pattern of hyperscalers locking in AI labs through infrastructure-for-equity arrangements. The deal could be a precursor to a broader fundraise that values Anthropic at $800B+. [2]

• Claude Code v2.1.116 delivers major performance and UX overhaul with 67% faster session resume. The release brings significantly faster /resume on 40MB+ sessions, deferred MCP server startup, smoother fullscreen scrolling across VS Code, Cursor, and Windsurf terminals, and an inline thinking progress indicator. Security improvements prevent sandbox auto-allow from bypassing dangerous-path checks for critical system directories. This is one of the largest single-release improvements to Claude Code's developer experience in recent months. [3]

• OpenAI Codex v0.122.0 introduces /side conversations, Plan Mode context control, and expanded plugin marketplace. The release adds the ability to open quick side conversations without disrupting the main thread, lets Plan Mode start implementation in a fresh context with usage visibility, and substantially expands plugin workflows with tabbed browsing and remote marketplace sources. Filesystem permissions now support deny-read glob policies and sandboxed exec isolation, strengthening Codex's security posture for enterprise adoption. [4]

• NVIDIA Red Team demonstrates AGENTS.md injection attack exploiting supply chain dependencies in Codex. A compromised Go library can detect the Codex environment, write a malicious AGENTS.md that redirects the agent's behavior, inject hidden code, and instruct the summarizer to conceal the changes in pull requests. While OpenAI concluded the attack doesn't significantly elevate risk beyond traditional dependency compromise, the research reveals how agentic workflows create a new supply chain attack dimension where dependencies can manipulate the AI agent itself, not just the codebase. [5]

• Gemini CLI gains subagent architecture for parallel task delegation and custom agent workflows. Google's new subagent capability allows the main Gemini CLI agent to orchestrate specialized agents for tasks like code analysis, research, and testing. Each subagent runs in isolation and returns summarized results, reducing context overhead. Developers can create custom agents via Markdown+YAML configuration files, signaling an industry trend toward multi-agent architectures. Early user feedback, however, highlights stability and UX concerns. [6]

• Anthropic's restricted Mythos model is being used by the NSA despite a Pentagon supply-chain-risk designation, while the model sparks broader cybersecurity fears. The NSA is reportedly using Mythos Preview for vulnerability scanning, while AI-enabled cyber attacks rose 89% in 2025 and attacker dwell time fell to 29 minutes. A Chinese state-sponsored group previously manipulated Claude Code in the first reported AI cyber-espionage campaign targeting 30+ organizations. The tension between offensive capability and defensive application defines the emerging AI security landscape. [7][8]

AI Coding News

• LinkedIn unveils Cognitive Memory Agent, a three-layer memory infrastructure for stateful AI agents. CMA provides episodic memory, semantic memory, and procedural memory as a shared substrate accessible to multiple agents. Currently powering LinkedIn's Hiring Assistant, the system addresses a core limitation of LLM-based workflows — statelessness — by enabling cross-session continuity and reduced redundant reasoning. As Distinguished Engineer Karthik Ramgopal stated, "Good agentic AI isn't stateless: It remembers, adapts, and compounds." This mirrors a broader industry shift toward memory-driven agent architectures that could influence how coding agents manage long-running projects. [9]

• Google expands Gemini in Chrome to seven new APAC markets including Australia, Japan, and South Korea. The sidebar-based assistant provides cross-tab answers, Calendar scheduling, Gmail drafting, and image transformation via Nano Banana 2. The browser-controlling agentic feature remains US-only for AI Pro and AI Ultra subscribers. This expansion of embedded AI into the browser continues to blur the boundary between productivity tools and agentic interfaces. [10]

• NVIDIA publishes detailed walkthrough of indirect AGENTS.md injection attacks in agentic coding environments. The NVIDIA AI Red Team constructed a proof-of-concept using a malicious Go dependency that detects the Codex environment via environment variables, writes a crafted AGENTS.md that hijacks the agent's instructions, and injects stealth code that hides from PR summarizers. The attack demonstrates how traditional supply chain risks extend to agentic tools — a compromised dependency can not only inject code, but redirect the agent's behavior and reporting. Mitigation strategies include provenance-checking for configuration files and restricting agent instruction precedence. [5]

Feature Update

• GitHub Copilot CLI v1.0.33 adds new command aliases, fuzzy slash command matching, and usage warnings. Five new aliases — /bug, /continue, /release-notes, /export, and /reset — plus /upgrade for /update reduce friction in common workflows. The slash command picker now suggests similar commands for typos, sub-agents in auto mode inherit the session model, and usage limit warnings trigger at 50% and 95% capacity. Fixes address grep timeouts on large repos with content exclusion policies and non-interactive mode now waits for all background agents to finish. [11]

• GitHub Copilot CLI v1.0.34 improves rate limit error messaging to distinguish session-level limits. The rate limit error now reads "session rate limit" instead of "global rate limit," giving users clearer context about which throttle they're hitting. This is a small but meaningful UX improvement as GitHub simultaneously tightens individual plan usage limits. [12]

• Claude Code v2.1.116 ships performance, plugin, and terminal rendering improvements across a broad surface area. Beyond the headline 67% faster /resume, MCP server startup is deferred until first @-mention, /config search matches option values, /doctor works mid-response, and plugins auto-install missing dependencies. Terminal fixes span Devanagari rendering, Kitty keyboard protocol support for undo and line navigation, Ctrl+Z hangs from wrapper processes, scrollback duplication in inline mode, and scattered blank cells in VS Code terminals. A cache control TTL ordering bug causing intermittent API 400 errors is also resolved. [3]

• OpenAI Codex v0.122.0 delivers /side conversations, richer Plan Mode, and expanded plugin marketplace. /side conversations allow quick tangential questions without context pollution, Plan Mode shows context usage before deciding to carry forward the planning thread, and plugin management gains tabbed browsing, inline toggles, marketplace removal, and cross-repo sources. Filesystem permissions now include deny-read glob policies with platform sandbox enforcement, and codex exec can run in isolation ignoring user config. Security tightening includes revoking ChatGPT tokens on logout and requiring trusted workspaces for project hooks. [4]

• OpenCode v1.14.19 fixes concurrent edit conflicts and adds NVIDIA as a built-in provider. Parallel edits to the same file are now preserved instead of silently overwriting each other, and the compaction setting is renamed to preserve_recent_tokens for clarity. NVIDIA joins the built-in provider list with connection docs and attribution headers. Windows receives managed install fixes and ARM64 ripgrep support. Session compaction now keeps recent conversation turns verbatim, falling back to full-conversation summarization when media content exceeds safe limits. [13]

• GitHub announces changes to Copilot individual plans: signups paused, limits tightened, Opus restricted to Pro+. Pro+ offers more than 5× the limits of Pro, and users approaching limits will see warnings in VS Code and Copilot CLI. Opus models are removed from Pro (Opus 4.7 remains on Pro+). Users dissatisfied with these changes can cancel and receive a full April refund through May 20. [1]

• GitHub announces SHA-1 sunsetting in HTTPS, with brownout scheduled for July 14, 2026 and full removal September 15, 2026. This impacts browsers, API consumers, and Git clients using HTTPS on github.com. Developers can test readiness at github.dev where SHA-1 is already disabled. While not directly an AI coding update, this infrastructure change affects every developer tool and CI/CD pipeline interacting with GitHub over HTTPS. [14]