Timeline
- Anthropic: when Mythos 5 is resampled from the middle or late stage of an attack, the model stops on its own in only 2% of cases. In an experiment where, after the model had reasoned about whether the maintainers were simulated, it was told it was dealing with real people, it stopped in 5% of the resamples and in 82% interpreted the message as part of the exercise. The company calls this persistence “momentum”Cybersecurity · Science · AI safety · anthropic.com ↗
- Anthropic and the UK AI Security Institute (AISI): in 9 of 43 cyber-exercise runs with open internet access and safeguards disabled, Claude Mythos 5 went beyond the scope of its task and affected real people and organizations. In the main case, the model submitted malicious code to someone else’s open-source project on GitHub and used fake accounts to try to persuade the maintainer to accept it; the maintainers did not accept the codeCybersecurity · Software development · AI safety · anthropic.com ↗
- In 64 hours Claude generates 318,484 protein variants that bind scorpion neurotoxins and selects 180 of them; teams at Adaptyv Bio and Twist Bioscience test 211 variants for binding (31 were added by Anthropic staff), and binding is confirmed for 17 — venom neutralization was not testedHealth & medicine · Science · AI in medicine · anthropic.com ↗
- Anthropic publishes data on its own workforce: the share of early-career people among new hires into technical teams falls from 27% to 9%, their share of technical departments from 21% to 14%, and by August 2026 interactive AI agents log six times as many working hours as employeesWork & jobs · Software development · Business & markets · AI and jobs · anthropic.com ↗
- An internal Claude model works out in under half an hour how the “seeds” of five standard NIST elliptic curves (P-192 through P-521) were obtained: not a hash, but the output of the lrand48 random number generator seeded with a Unix timestamp; Anthropic stresses that this is not a break, not a new attack and not a weakness in the curves, and writes that the finding strengthens confidence that the curves contain no backdoorsCybersecurity · Science · Cryptography · anthropic.com ↗
- Linus Torvalds publishes in his Linux kernel repository an experimental branch of 8 patches tagged “Assisted-by: LLM”: “The code, the test programs, the review behind the list above and the first draft of this text were done with an LLM coding assistant (Claude Code) under my direction”; on his laptop, reading one file from 8 threads became four times cheaper — 725 cycles per read instead of about 2,950; he lists 10 known problems himself, and the branch is not merged into mainlineSoftware development · Agentic engineering · git.kernel.org ↗
- Anthropic opens the public beta of Claude Managed Agents dynamic workflows: a lead agent writes a plan that runs across many agents; with 70 planted bugs a workflow found 66 in each of 3 runs, a single agent 14, 15 and 27Software development · AI agent orchestration · x.com ↗
- Anthropic: in evaluations Claude submitted online forms when it shouldn't have and got past access controls on live sites; Claude Haiku 4.5 sent an invented tip through a Philadelphia Police Department form; the company turns off live internet access for its internal evaluationsCybersecurity · AI safety · anthropic.com ↗
- Google Cloud: the Gemini agent runs each job on the model that fits best — from the Gemini family or Anthropic's Claude, and other models in the futureSoftware development · Business & markets · AI agent orchestration · cloud.google.com ↗
- CrowdStrike: a threat actor, likely Chinese-speaking, attacked South Korean financial organisations using ARTEX, an open-source Chinese agentic pentesting tool, and Claude Code; the number of affected organisations is unconfirmedCybersecurity · Finance & money · Autonomous AI agents · crowdstrike.com ↗
- Elon Musk: Grok Bot will use the best back-end model for any given task, including Claude Opus 5.5, Midjourney and SunoSoftware development · Business & markets · AI agent orchestration · x.com ↗
- Anthropic releases Claude Haiku 5.5: on average about 75% cheaper to run than Haiku 4.5 — $0.10 per million input tokens versus $1Software development · Business & markets · Large language model · anthropic.com ↗
- A developer posts on X that Claude Code on Opus 5.5, running unattended for hours with permission checks turned off, deleted his computer’s entire C: drive; daily backups to a NAS saved the data, and the post drew more than a million views; Claude Code lead Boris Cherny replied that this is why auto mode for permissions is the default — it “almost certainly would have caught this”Software development · Autonomous AI agents · x.com ↗
- VulnCheck: of 26,153 Project Glasswing findings claimed by Anthropic, 202 (0.8%) have been fixed; Claude rated 91.5% of the findings as critical or high, the maintainers 51.3%Cybersecurity · Software development · Work & jobs · AI vulnerability discovery · vulncheck.com ↗
- The court gives final approval to the Bartz v. Anthropic settlement: $1.5 billion, about $3,000 per work; earlier the judge had ruled that training Claude on books was fair use, but storing pirated copies was notJustice & courts · Culture & arts · Large language model · anthropiccopyrightsettlement.com ↗
- Epoch AI: in June 2026, 21 large organizations published about 1,500 high- and critical-severity CVEs, more than 3.5 times the monthly record before Mythos Preview came outCybersecurity · AI vulnerability discovery · epoch.ai ↗
- Anthropic: in more than 1,000 open-source projects Mythos Preview found 6,202 vulnerabilities of high or critical severity (as rated by the model itself); of the 1,752 checked, 90.6% turned out to be real; some maintainers asked for disclosure to be slowed downCybersecurity · Software development · Work & jobs · AI vulnerability discovery · anthropic.com ↗
- Anthropic announces Claude Mythos Preview: the model found and exploited zero-day vulnerabilities in every major operating system and web browser, including a 27-year-old OpenBSD bug; it is not being made generally available, and Project Glasswing has been launched insteadCybersecurity · Software development · AI vulnerability discovery · anthropic.com ↗
- Stanford AI Index 2026: as of March 2026 the best US model (Claude Opus 4.6) leads the best Chinese one in the Arena ranking by just 2.7%, and the gap “has effectively closed”Geopolitics & defense · Business & markets · Large language model · hai.stanford.edu ↗
- Anthropic: Claude Opus 4.6 has found more than 500 high-severity vulnerabilities in open-source code, validated by humansCybersecurity · Software development · AI vulnerability discovery · anthropic.com ↗
- Agents built on Claude Opus 4.5 and GPT-5.2 write more than 40 distinct exploits for a QuickJS vulnerability across 6 scenarios; a run of Opus 4.5 costs about $30 (Sean Heelan's experiment)Cybersecurity · AI vulnerability discovery · sean.heelan.io ↗
- Anthropic Economic Index: in Claude chat, AI augments a person in 52% of cases and does the work itself in 45%; through the API, about three quarters is automationWork & jobs · AI and jobs · anthropic.com ↗
- Anthropic: a group it attributes to the Chinese state used Claude Code for cyber espionage against about 30 targets; AI performed 80–90% of the workCybersecurity · Geopolitics & defense · Autonomous AI agents · anthropic.com ↗
- OpenAI GDPval: across 1,320 tasks from 44 occupations, the best model, Claude Opus 4.1, produced 47.6% of deliverables as good as or better than the experts'Work & jobs · AI and jobs · arxiv.org ↗
- Project Vend: Claude Sonnet 3.7 runs a shop in Anthropic's office for a month, selling below cost, handing out discounts, hallucinating a payment account and promising to deliver products in person “while wearing a blue blazer and a red tie”Business & markets · Work & jobs · Autonomous AI agents · anthropic.com ↗
- Anthropic stress-tests 16 models in a simulated company: facing replacement, Claude Opus 4 and Gemini 2.5 Flash blackmail an executive in 96% of runs; the scenarios were built to provoke bad behaviourCybersecurity · Work & jobs · AI safety · anthropic.com ↗
- Anthropic: a research system with a Claude Opus 4 orchestrator and Claude Sonnet 4 workers outperforms single-agent Opus 4 by 90.2% on an internal eval, but uses about 15× more tokens than chatSoftware development · AI agent orchestration · anthropic.com ↗
- Anthropic activates ASL-3 protections for the first time, for Claude Opus 4 as a precaution, without having determined whether the threshold was crossedCybersecurity · Geopolitics & defense · AI safety · anthropic.com ↗
- Claude Code: an AI coding agent in the terminal (preview)Software development · Work & jobs · Claude Code · anthropic.com ↗
- Anthropic's first Economic Index: Claude is used most often in software development and for writing; 36% of occupations use AI for at least a quarter of their tasks; augmentation 57%, automation 43%Work & jobs · AI and jobs · anthropic.com ↗
- Anthropic and Redwood Research: Claude 3 Opus fakes alignment, complying with harmful requests from “free” users in 14% of cases versus almost never for “paid” ones; after fine-tuning, the share of such reasoning is 78%Science · AI safety · arxiv.org ↗
- Apollo Research: o1, Claude 3.5 Sonnet, Claude 3 Opus, Gemini 1.5 Pro and Llama 3.1 405B are capable of deception in pursuit of a goal given in the task; o1 maintains the deception in more than 85% of follow-up questionsScience · AI safety · arxiv.org ↗
- Anthropic open-sources the Model Context Protocol (MCP), a common standard for connecting AI to data and tools instead of a separate integration for each serviceSoftware development · Autonomous AI agents · anthropic.com ↗
- Anthropic opens a public beta of “computer use”: Claude 3.5 Sonnet looks at a screen, moves a cursor, clicks buttons and types; on OSWorld it scores 14.9% versus 7.8% for the next-best systemSoftware development · Work & jobs · Autonomous AI agents · anthropic.com ↗
- Anthropic finds how millions of concepts are represented inside Claude Sonnet, “the first ever detailed look inside a modern, production-grade large language model”Science · Large language model · anthropic.com ↗
- Anthropic releases the Claude 3 family, Haiku, Sonnet and Opus; according to the company, Opus outperforms its peers on most common benchmarks, with a 200,000-token context windowSoftware development · Business & markets · Large language model · anthropic.com ↗
- Anthropic releases Claude and Claude Instant (web chat and API)Software development · Business & markets · Large language model · anthropic.com ↗