Anvat / News

AI + security news that matters.

Curated takes on Anthropic / OpenAI / Google releases and AI-relevant security events — short, opinionated, deep-dive links where it counts.

  1. OpenAI#

    GPT-5.6 'iris-alpha' surfaces in OpenAI Codex logs — Polymarket prices June 30 at 89%

    The identifier gpt-5.6 appeared in OpenAI's Codex backend logs on May 1 before disappearing. Codenames iris-alpha, kindle-alpha, kepler-alpha have leaked. Polymarket is pricing 80-89% odds for a public release by June 30, 2026. A 1.5M-token context window is rumored but unconfirmed.

    Full release tracker
  2. Policy#

    EU AI Act enforcement powers activate August 2, 2026 — 57 days to prepare

    The European Commission's enforcement powers over general-purpose AI (GPAI) model providers activate August 2, 2026. Fines up to 3% of global annual turnover. The Code of Practice finalized in June 2026; signatories get a one-year good-faith enforcement honeymoon. Builders consuming GPAI through APIs are affected via Chapter III obligations on AI Systems.

    Builder's guide
  3. OpenAI#

    Cursor SDK adds custom stores, custom tools, and nested subagents

    TypeScript and Python SDKs gain JSONL agent stores (version-controllable), bring-your-own functions exposed as model tools, auto-review routing for local tool calls, and arbitrary subagent nesting. Aimed at production scripts, CI pipelines, and custom integrations.

    Cursor vs Antigravity vs Devin
  4. Crypto#

    Zcash NU6.2 hard fork activates corrected Orchard circuit

    Five days after a critical soundness vulnerability was discovered in the Orchard zero-knowledge proof circuit, the NU6.2 hard fork activated successfully and re-enabled shielded transactions. The bug had been live since the pool launched in May 2022. Discovered using Claude Opus 4.8.

    Full timeline + workflow
  5. Claude#

    Claude 'Oceanus' model identifier appears in Claude Console

    A new model identifier — claude-oceanus-v1-p — appeared in the Claude Console with red-team access granted around the same day. Oceanus appears to be the next preview candidate in the Mythos line, focused on agentic coding + offensive security rather than chat. Red-team-to-launch lag has been 1-2 weeks historically — possible public launch second half of June.

    What Mythos is
  6. OpenAI#

    Cognition AI rebrands Windsurf as Devin Desktop; Cascade retires July 1

    Six months after the $250M acquisition, Cognition consolidated Windsurf under the Devin brand. The IDE shell stays, the in-house Cascade agent retires July 1, 2026, and the new flow hands off cloud work to Devin. SWE-1.6 in-house model hits 950 tok/s on the new Pro+ tier ($200/mo).

    IDE comparison
  7. Security#

    Security researcher finds 4-year-old Zcash ZK proof bug with Claude Opus 4.8

    Independent researcher Taylor Hornby discovered a critical soundness bug in Zcash's Orchard circuit that allowed unlimited counterfeit ZEC minting inside the shielded pool. The exploit was written using Anthropic's publicly available Opus 4.8 — not the Glasswing-gated Mythos. ZEC dropped 31% intraday before partial recovery.

    Read the workflow
  8. Claude#

    Anthropic expands Project Glasswing to 150 new organizations

    Glasswing — the program gating Mythos Preview access — added ~150 new vetted organizations across 15+ countries, including reported additions Okta, Samsung, ENISA, and NATO. Partners have surfaced 23,000+ candidate vulnerabilities, 6,000+ confirmed severe, since the April launch. General availability of Mythos-class capability committed to but no timeline.

    What Glasswing means
  9. Claude#

    Claude Security launches in public Enterprise beta — runs on Opus 4.7

    Anthropic's Claude Security product entered public Enterprise beta, reporting 2,100+ vulnerabilities patched in three weeks. Important detail: the production product runs on Opus 4.7, not Mythos. The 'Mythos 1' toggle that briefly appeared in the UI is a placeholder for future Mythos-class capability — not active production.

  10. Google#

    Google ships Gemini 3.5 Flash — beats Gemini 3.1 Pro on agentic + coding

    Gemini 3.5 Flash went GA at I/O 2026 — Terminal-Bench 76.2%, MCP Atlas 83.6%, GDPval-AA Elo 1656, all ahead of last-gen Gemini 3.1 Pro. Regressed on Humanity's Last Exam, ARC-AGI-2, and 128K long-context recall. Pichai confirmed Gemini 3.5 Pro ships in June.

    Gemini 3.5 Pro tracker
  11. Google#

    Google launches Antigravity 2.0 — Browser Subagent + parallel agent orchestration

    Google's AI IDE refreshed at I/O 2026 with a headless Chromium 'Browser Subagent' that QAs your app as you build, plus a Go-based CLI and developer SDK. Gemini 3.5 Flash is the default at 289 tokens/sec. $20/mo Pro tier; quota lockouts hit Pro subscribers within the first week before limits were raised.

    Full IDE comparison
  12. OpenAI#

    DeepSeek releases V4 — 1.6T MoE, 1M context, 80.6% SWE-Bench Verified

    DeepSeek-V4-Pro (1.6T / 49B active) and V4-Flash (284B / 13B active) shipped MIT-licensed with 1M-token context. V4-Pro hits 80.6% on SWE-Bench Verified (within 0.2pt of Opus 4.6) and 93.5 on LiveCodeBench (highest of any model). Pricing: $1.74/$3.48 per MTok — about 1/20th of Opus 4.7's output rate.

    V4 Pro deep-dive
  13. Claude#

    Anthropic announces Claude Mythos Preview — gated behind Project Glasswing

    Anthropic's most capable model to date — 93.9% on SWE-bench Verified, 83.1% on CyberGym vulnerability reproduction — launched as a gated research preview rather than a generally available product. 12 launch partners include AWS, Apple, Microsoft, Google, Cisco, CrowdStrike, NVIDIA, JPMorgan, and the Linux Foundation. $25/$125 per MTok partner pricing.

    Full explainer
  14. OpenAI#

    Cursor 3.0 ships — Agents Window, /worktree and /best-of-n commands

    Cursor 3.0 reframed the IDE around the Agents Window — run many agents in parallel across local repos, git worktrees, cloud VMs, remote SSH. New /best-of-n command runs the same task across multiple models in isolated worktrees and compares outcomes. Cursor 3.5 (May 2026) added 8-way parallel cloud agent execution.

    vs Antigravity + Devin
  15. Claude#

    Anthropic releases Agent Skills as open standard — adopted by Cursor, OpenAI, Google in 48h

    Claude Skills (shipped October 2025) released as a SKILL.md open standard at agentskills.io with governance donated to the Agentic AI Foundation under the Linux Foundation. By March 2026, 32+ AI tools adopted the format — including Microsoft VS Code, OpenAI Codex CLI, Google Gemini CLI, JetBrains Junie, AWS Kiro, Cursor.

    Full Skills guide