AI News Of The Week (20th February, 2026)
TLDR: Key AI Developments This Week
Anthropic launched Claude Sonnet 4.6 with a 1M context window. Google released Gemini 3.1 Pro and Lyria 3 music generation. Anthropic also introduced Claude Code Security. India operationalized new synthetic media rules. An AWS incident highlighted the risks of agentic coding tools.
Anthropic Launches Claude Sonnet 4.6 with Long-Context and Stronger Computer Use
Anthropic released Claude Sonnet 4.6 as a full upgrade across coding, computer use, long-context reasoning, agent planning, and knowledge work, featuring a 1M-token context window in beta. Sonnet 4.6 became the new default for Free and Pro plans at $3/$15 per million tokens — unchanged from Sonnet 4.5. A key security detail: the release explicitly addresses prompt injection resistance for computer-use agents, with Anthropic claiming a major improvement over Sonnet 4.5 and performance comparable to Opus 4.6.
Sonnet 4.6 Rolls Into GitHub Copilot and Amazon Bedrock on Launch Day
On the same day as its release, Claude Sonnet 4.6 began rolling into major developer platforms. GitHub added Sonnet 4.6 to Copilot across VS Code, Visual Studio, github.com, mobile, CLI, and the Copilot Coding Agent for Pro, Pro+, Business, and Enterprise users, with a 1x premium request multiplier and gradual rollout. AWS also made Sonnet 4.6 available in Amazon Bedrock, positioning it as a direct upgrade to Sonnet 4.5 and highlighting context-compaction capabilities. The same-day multi-platform landing signals a shrinking gap between model announcement and enterprise procurement.
Google Adds Lyria 3 Music Generation to Gemini, with SynthID Watermarking
Google rolled out Lyria 3 music generation inside Gemini, enabling users to create 30-second tracks from text prompts or uploaded photos and videos, with automatic lyric generation if desired. The release brings improved creative controls for style, vocals, and tempo. All generated tracks are embedded with SynthID, Google's imperceptible audio watermark, and users can upload audio to check for SynthID presence. Google states prompts naming specific artists are treated only as broad inspiration, with output filters applied to check against existing content.
Google Releases Gemini 3.1 Pro, Pairing Benchmark Narrative with a Detailed Model Card
Google announced Gemini 3.1 Pro as an upgraded core intelligence model available in preview across the Gemini API (including Google AI Studio and Gemini CLI), Vertex AI, Gemini Enterprise, the Gemini app, and NotebookLM. The model supports multimodal inputs (text, images, audio, video) with a 1M-token context window and 64k-token output limit. Google's headline benchmark is a verified 77.1% on ARC-AGI-2, described as more than double the reasoning performance of Gemini 3 Pro. The accompanying model card includes frontier safety discussion and cyber capability evaluation notes. Gemini 3.1 Pro also rolled into GitHub Copilot's model picker on the same day.
Anthropic Launches Claude Code Security and Open-Sources a Security Review Action
Anthropic announced Claude Code Security as a limited research preview built into Claude Code on the web, designed to scan codebases for vulnerabilities and propose targeted patches for human review. Unlike rule-based static analysis, it uses semantic reasoning over component interactions, data flow, and business logic. The system runs multi-stage verification — attempting to disprove its own findings — and applies severity and confidence ratings, with nothing applied without explicit human approval. Anthropic openly acknowledged dual-use risk and limited the preview to Enterprise and Team customers. Separately, Anthropic open-sourced a related GitHub Action for PR diff-aware security scanning, noting in its own documentation that it is "not hardened against prompt injection attacks" and should be used only for trusted PRs.
India's Synthetic-Media Rules Take Effect, Mandating Labels and Provenance Signals
India's Ministry of Electronics and Information Technology (MeitY) operationalized IT Rules amendments defining "synthetically generated information" (SGI) as AI-generated or AI-altered audio/visual content made to appear real and likely indistinguishable from genuine persons or events. Platforms are now required to add prominent labels and embed permanent provenance metadata — with anti-tampering obligations — and significant social media intermediaries must collect user declarations about SGI and apply technical verification before publishing. The rules also establish a three-hour takedown window once "actual knowledge" arises via court order or government intimation.
AI Coding Agents and Real Outages: AWS Incident Highlights Permission and Approval Fragility
Reporting this week described how an AI coding tool (Kiro) contributed to an AWS service disruption in mainland China in December by deleting and recreating an environment. AWS characterized the event as user error affecting one service in one region, while the Financial Times described internal concerns about reliability and oversight. The practical lesson for engineering teams deploying agentic tools is not about AI autonomy in the abstract but about control-plane design: human approvals, least-privilege, and rollout gating can fail in ways that allow highly capable tools to execute destructive actions rapidly — aligning with the broader week-spanning theme of vendors shipping agentic capabilities while simultaneously building guardrails.
