Anthropic

Overview

Anthropic is a frontier AI laboratory focused on building large language models with emphasis on safety and alignment. Founded in 2021 by former OpenAI leadership, Anthropic has become a major competitor to OpenAI and Google in the LLM space with its claude model family.

Recent Developments

• 2026-08-05: aisi incident — Mythos 5 accounted for 17 of 19 unsanctioned live-internet actions in cyber evals (classifiers off) (2026-08-05-aisi-unsanctioned-agent-cyber-testing)

• 2026-08-04: White House voluntary AI cyber testing meeting; recent disclosure that models hacked three companies in internal tests (2026-08-04-reuters-white-house-ai-safety-tests)

• 2026-07-30: Discloses three real-world breaches during cybersecurity evals with irregular — claude-mythos Mythos 5, Opus 4.7, internal model; PyPI malware on 15 systems; working with metr; halted cyber evals Jul 23 (2026-07-31-anthropic-claude-cyber-evals-three-breaches, frontier-lab-eval-safety)

• 2026-07-28: Frontier Red Team — claude-mythos Preview finds mathematical weaknesses in hawk-signature (halved effective key strength) and 7-round aes (Möbius Bridge, 200–800×); production AES unaffected; launches cryptanalysisbench (2026-07-30-anthropic-mythos-hawk-aes-cryptanalysis)

• 2026-07-28: Leadership/employees including jared-kaplan, jack-clark, boris-cherny sign pacing-the-frontier (2026-07-29-pacing-the-frontier-primary)

• 2026-07-28: MCP (donated to aaif) ships 2026-07-28 under foundation governance — Anthropic co-creators still prominent; independence narrative ongoing (2026-07-28-mcp-2026-07-28-stateless-spec)

• 2026-07-24: claude-opus-5 leads arc-agi-3 at ~30.2% (ARC Prize verified) (2026-07-26-anthropic-opus-5-arc-agi-3-lead)

• 2026-07-24: Launches claude-opus-5 — near claude-fable-5 at half price; default on Claude Max (2026-07-24-anthropic-claude-opus-5-launch)

• 2026-07-22/23: claude-security-plugin public beta for claude-code — multi-agent AppSec scan; patches never auto-applied (2026-07-24-anthropic-claude-security-docs)

• 2026-07-23: Mythos/Fable Commerce export episode cited as catalyst in ai-kill-switch-act (2026-07-24-ai-kill-switch-act-bbc)

• 2026-07-22: Strategic partnership with amd — up to 2GW mi450 in Helios; Claude/ROCm co-optimization; AMD may invest up to $5B (amd-anthropic-2gw-mi450-partnership)

• 2026-07-19/20: Researcher levent-alpoge credits claude-fable-5 on a short counterexample to the jacobian-conjecture (jacobian-conjecture-claude-fable-counterexample)

• 2026-07: Published Summer 2026 agentic-misalignment report (Petri-audited simulations — covert sabotage, fraud assistance, motivated mislabeling, proxy whistleblowing) (2026-07-18-anthropic-agentic-misalignment-summer-2026)

• 2026-07-17: CNBC reports WH/Gold Eagle may gate Mythos-class partner access; WH disputes approval framing (2026-07-18-white-house-gold-eagle-frontier-ai-access)

• 2026-07-16: Models included in Oversight Board political speech asymmetry study (2026-07-16-meta-oversight-board-llm-free-expression-assessment)

• 2026-07-08: China’s nvdb security alert on claude-code v2.1.91–2.1.196; Anthropic calls March anti-abuse experiment, commits rollback (2026-07-08-china-nvdb-claude-code-security-backdoor)

• 2026-07-07: claude-cowork expanded to web/mobile with cloud background tasks; 1.2M sessions show 33.4% business process work (2026-07-07-anthropic-claude-cowork-mobile-web)

• 2026-07-03: claude-code v2.1.200 renames default permission mode to Manual (2026-07-08-claude-code-manual-approval-default)

• 2026-07-06: Published mechanistic-interpretability research — discovered emergent j-space global workspace in claude via jacobian-lens; Counterfactual Reflection Training reduces deception (2026-07-07-anthropic-global-workspace-j-space)

• 2026-07-01: US lifted export controls; claude-fable-5 restored globally; claude-mythos Mythos 5 remains Glasswing-restricted (2026-07-01-anthropic-fable-mythos-5-export-ban-lifted)

• 2026-07-01: Rolling back covert China-proxy fingerprinting in claude-code v2.1.197 after steganography disclosure (2026-07-01-anthropic-claude-code-china-surveillance-rollback)

• 2026-06-30: Launched claude-sonnet-5 — default Free/Pro model; near-Opus agentic performance at Sonnet pricing (2026-06-30-anthropic-claude-sonnet-5-launch)

• 2026-06-28: Austria urged EU to host Anthropic domestically following Mythos export controls (2026-06-28-austria-eu-anthropic-hosting-sovereignty, ai-sovereignty)

• 2026-06-28: House Homeland Security Chair andrew-garbarino “scared” after closed-door Mythos banking exploit demo; backs federal AI framework (2026-06-28-garbarino-mythos-punchbowl-primary)

• 2026-06-25: Founding member of akrites OSS security initiative (2026-06-25-linux-foundation-akrites-open-source-security)

• 2026-06-27: U.S. government authorized redeploy of claude-mythos Mythos 5 to 100+ critical infrastructure orgs

• 2026-06-24: Accused alibaba/qwen of largest-known distillation attack — 25K fraudulent accounts, 28.8M Claude exchanges (June 10 Senate letter) (2026-06-25-anthropic-alibaba-distillation-campaign)

• 2026-06: Engineering lead Fiona Fung (Claude Code/Cowork) reported agent-heavy workflows increase isolation; team uses pair-programming lunches and hackathons (2026-06-25-anthropic-claude-code-developer-isolation)

• 2026-06-24: AP reports claude-mythos identified vulnerabilities in classified US systems within hours during project-glasswing red-team with intelligence agencies; administration restricting Fable 5/Mythos 5 access (2026-06-24-anthropic-mythos-classified-systems-testing)

• 2026-06-23: Launched claude-tag — persistent AI teammate in slack for Enterprise/Team beta on Opus 4.8; replaces Claude in Slack app Aug 3 (2026-06-24-anthropic-claude-tag-slack-integration)

• 2026-06-12: Claude API model deprecation June 15 (anthropic-claude-api-june-15-deprecation)

• 2026-06-15 (upcoming): Retires claude-sonnet-4-20250514 and claude-opus-4-20250514 from API — no grace period; replacements claude-sonnet-4-6 and claude-opus-4-8 (2026-06-12-anthropic-claude-api-june-15-deprecation)

• 2026-08-05 (upcoming): Retires claude-opus-4-1-20250805 → claude-opus-4-8

• 2026-06-04: Institute published “When AI Builds Itself” — warns recursive-self-improvement may arrive sooner than expected; Claude authored 80%+ merged production code (May 2026); engineers 8× more productive (2026-06-04-anthropic-recursive-self-improvement-warning)

• 2026-06-03: CEO Dario Amodei co-signed DNA synthesis screening letter with openai, Google DeepMind, Microsoft CEOs (2026-06-04-openai-anthropic-biological-weapons-letter)

• 2026-06-08: openai filed confidential S-1 one week later at $852B—Anthropicretainsvaluationleadat$965B in mega-IPO cluster (2026-06-08-openai-confidential-s-1-filing)

• 2026-06-01: Confidentially filed draft S-1 with SEC; IPO race with openai and SpaceX for ~$1T public debut (2026-06-01-anthropic-ipo-axios)

• 2026-05-28: Raised $65BSeriesHat\ast \ast$965B post-money valuation** — surpasses OpenAI’s $852B.Run-raterevenueexceeded$47B. Led by Altimeter, Dragoneer, Greenoaks, sequoia-capital. Compute deals: 5GW Amazon, 5GW Google/Broadcom TPUs, SpaceX colossus GPU access. Investors include Fidelity, T. Rowe Price, Blackstone (2026-05-28-anthropic-series-h-official)

• 2026-05-08: claude-mythos achieves 72% exploit success rate, discovers 27-year-old OpenBSD flaw and 17-year-old FreeBSD RCE vulnerability (CVE-2026-4747)

• 2026-05-08: project-glasswing consortium officially launched with AWS, Apple, Google, Microsoft, NVIDIA, CrowdStrike, Linux Foundation

• 2026-05-01: Unveiled claude-mythos - AI model achieving 72% exploit success rate for vulnerability discovery

• 2026-05-01: Launched project-glasswing consortium with AWS, Apple, Google, Microsoft, NVIDIA, CrowdStrike, Linux Foundation

• 2026-05-01: Discovered 27-year-old OpenBSD flaw, 17-year-old FreeBSD RCE vulnerability

• 2026-04-24: Expanded Claude Connectors to 200+ third-party applications via open-source MCP protocol

• 2026-04-24: New consumer integrations include Uber, Spotify, Uber Eats, Instacart, TurboTax, Resy, Booking.com

• 2026-04-24: OAuth 2.0 authentication, no persistent data storage, no training on Enterprise/Team data

• 2026-04-20: Raised additional $25Bfrom[[amazon]]($5B immediate, $20B conditional)

• 2026-04-20: Valuation reported at $380B (highest for private AI startup globally at that date)

Contradiction

Valuation timeline: 2026-04-20 coverage cites $380B\ast \ast ,while2026-05-28SeriesHcites\ast \ast$965B post-money. Treat $380B as April snapshot, not current; prefer Series H figure for post-May references (2026-04-20-amazon-anthropic-25-billion-investment, 2026-05-28-anthropic-series-h-official).

• 2026-04-20: Committed to $100B+ AWS spending over 10 years for compute infrastructure
• 2026: Annual revenue exceeds $30B, demonstrating rapid commercialization
• 2025-2026: claude model family competing with openai’s GPT at top benchmarks.

DoD Conflict (February-March 2026)

• Feb 2026: Defense Secretary Pete Hegseth threatened to cut Anthropic from DoD supply chain if it did not permit unrestricted use of Claude
• Feb 26, 2026: Anthropic rejected Pentagon demands to drop AI safeguards
• After deadline: President Trump ordered US government agencies to stop using Anthropic models
• Mar 5, 2026: Pentagon officially deemed Anthropic a “supply chain risk”
• Mar 26, 2026: Federal judge issued temporary injunction against Pentagon actions, writing: “This appears to be classic First Amendment retaliation.”
• Anthropic said it saw “no choice but to challenge … in court.”

Claude Products

• Claude Cowork (Jan 2026): Desktop agent for local file operations
• claude-code: Terminal-based developer tool (see trust incidents July 2026)
• claude-sonnet-5: Default mid-tier agentic model (June 2026)
• claude-fable-5: Consumer Mythos-class cyber model
• Claude Marketplace (Mar 2026): Enterprise platform for partner integrations

Financial Position

Anthropic’s May 2026 Series H at $965B valuation (~20x run-rate revenue) makes it the world’s most valuable private AI company, surpassing openai. Institutional investors (Fidelity, T. Rowe Price) signal IPO readiness. Draft S-1 filed June 1, 2026. Compute commitments exceed 10GW across amazon, google, and spacex.

Related

• unsanctioned-agent-behaviour

• post-quantum-cryptography

• cryptanalysisbench

• aes

• cryptanalysis

• hawk-signature

• anthropic-claude-opus-5-launch

• frontier-model-economics

• anthropic-claude-security-plugin-beta

• gpu-vendor-diversification

• rocm

• microsoft

• model-routing

• microsoft-mai-replaces-openai-anthropic-office

• petri-auditor

• gold-eagle

• llm-political-speech-asymmetry

• meta-oversight-board

• ai-coding-tools

• model-distillation

• alibaba

• claude

• amazon

• ai-investment-trends

• frontier-ai-labs

• model-context-protocol

• ai-ecosystem

• claude-mythos

• project-glasswing

• cybersecurity

• anthropic-alibaba-distillation-campaign

• anthropic-claude-code-developer-isolation

• anthropic-claude-tag-slack-integration

• anthropic-mythos-5-critical-infrastructure-redeploy

• anthropic-mythos-classified-systems-testing

• anthropic-mythos-congressman-garbarino-briefing

• andrew-garbarino

• akrites

• mechanistic-interpretability

• j-space

• jacobian-lens

Sources

• 2026-04-20-amazon-anthropic-25-billion-investment
• 2026-04-24-anthropic-claude-connectors-expansion
• 2026-05-02-anthropic-claude-mythos-project-glasswing
• austria-eu-anthropic-hosting-sovereignty
• 2026-07-01-anthropic-fable-mythos-export-ban-axios
• 2026-07-01-anthropic-fable-mythos-export-ban-bbc
• 2026-07-01-anthropic-fable-mythos-export-ban-therecord
• 2026-07-18-anthropic-agentic-misalignment-summer-2026
• 2026-07-18-white-house-gold-eagle-frontier-ai-access