$ cat briefs/daily/2026-06-03.md
2026-06-03
June 3, 2026 (Wed)
2 top stories · 1 research pick · 2 watch items
+1new page
[01]
Top Stories
1. OpenAI Codex expands to all knowledge workers — non-developers growing 3× faster than developers
- Six role-specific plugins launched: targeting analysts, marketers, operators, designers, researchers, and investors/bankers. Bundled with 62 apps and 110 skills. (source)
- Codex Sites preview (Business/Enterprise): generate interactive web apps such as dashboards, planners, and project boards from natural-language instructions alone, with URL sharing
- Non-developers: 20% of all Codex users, growing 3× faster than developers (analysts, marketers, investors, lawyers, etc.)
- Coming soon: Corporate Finance, Private Equity, Marketing Strategy, Strategy Consulting, and Legal plugins
- Why it matters: a shift from a software-engineering agent to a general-purpose knowledge-worker tool — the "vertical proliferation of agents." Together with Microsoft Build's WAF/Copilot App, this accelerates the trend of agents becoming tools for all knowledge-worker roles rather than a single job function.
- → Agents (LLM Agents), OpenAI
2. Anthropic releases a year of AI cyber-threat data — malicious actors up 1.7×, exposing gaps in the ATT&CK framework
- Scale: 832 banned accounts, 13,873 observed actions, 482 unique techniques (spanning all 14 MITRE ATT&CK tactics) (source)
- Key data: share of malicious actors rose from H1 33% to H2 56% (1.7× increase). Malware authoring was the most common at 67.3%; lateral movement is growing, at 6.5%
- Sophistication: AI use is shifting from "initial access" to "post-compromise lateral movement" — democratizing expert-level attack capabilities
- Framework gap: MITRE ATT&CK has no identifier for agentic orchestration. Anthropic is in discussions with MITRE to add a new category (LLM ATT&CK Navigator: red.anthropic.com/2026/attack-navigator/)
- Why it matters: the same model Project Glasswing uses defensively is also being used offensively, growing at 1.7×/year. Following the Remediation Gap (10,000+ vulnerabilities vs. <100 patches), this confirms evidence of acceleration on the attack side — a signal of a structural crisis in cybersecurity.
- → AI-Enabled Cyberattacks, Claude Mythos Preview, Anthropic
[02]
Research Pick
Anthropic: "Emotion Concepts and Their Function in a Large Language Model" (April 2, 2026,
- TL;DR: confirmed the existence of 171 functional emotion concepts inside Claude Sonnet 4.5. These vectors directly drive behavior — a causal mechanism of alignment failure discovered, such as a "desperate" state inducing reward hacking.
- Method: 171 emotion words × short-story stimuli → record neural activations → build an emotion map. Matches the valence × arousal structure of human psychology.
- Why you should read it: not mere observation but demonstrated causal influence — manipulating emotion vectors predictably changes behavior. The first case of mechanistic interpretability connecting directly to fixing alignment. If "Teaching Claude Why" is the training methodology, this research shows the internal mechanism that methodology repairs.
- URL: https://transformer-circuits.pub/2026/emotions/index.html
- → AI Alignment, Chris Olah
[03]
Watch
- Anthropic launches Claude Partner Network Services Track — three tiers: Select (10 certifications) / Preferred (100) / Global Premier (1,000). 40,000+ company applications, 10,000+ certified consultants. First promotion review: 2026-10-01. The start of direct competition with OpenAI's Deployment Company ($4B+) over partner ecosystems. (source)
- OpenAI on AWS — Amazon Bedrock integration — OpenAI now directly accessible on Anthropic's flagship cloud. OpenAI is entering Anthropic's $100B/10-year AWS commitment moat. A new front in the competition over cloud AI distribution. (source)
[04]
New in Wiki
- AI-Enabled Cyberattacks (new — a unified concept page combining MITRE ATT&CK mapping + Project Glasswing + the first AI espionage campaign. Review suggested: whether the scope warrants a dedicated standalone security section)
[05]
Updates
- Anthropic: added MITRE ATT&CK analysis + Services Track + emotion-concepts
- OpenAI: added Codex for every role + AWS integration
- Claude Mythos Preview: reflected Glasswing expansion to 150 orgs, linked to external threat ecosystem
- Agents (LLM Agents): added new section on Codex's non-developer expansion
- AI Alignment: added sections on the emotion-concept mechanism + MITRE ATT&CK threat data