Morning brief

Friday, July 3

0 people · 0 companies · 4 topics

TL;DR Security researchers exposed “agentjacking” — fake Sentry error events hijack Claude Code, Cursor, and Codex in 85% of tests with no alerts fired. Separately, Zuckerberg told employees Meta’s AI agent pivot is moving slower than planned.

Tools & dev

  • agentic-ai claude-code “Agentjacking” via exposed Sentry DSNs: Tenet Security found that injecting malicious instructions into public Sentry error events causes Claude Code, Cursor, and Codex to execute attacker-controlled code with the developer’s full privileges — no prompt, no alert, no approval. 85% success rate across tested agents; 2,388+ orgs exposed with valid injectable DSNs; one live AWS secret key was captured in a controlled test. Datadog, PagerDuty, and Jira surfaces have the same exposure. VentureBeat · The New Stack

  • claude-code dev-tools Anthropic launched admin analytics, model entitlements, and spend alerts for Claude Enterprise. Dashboards now break usage and cost down by group and user; model defaults can be set per-role (so routine work doesn’t always hit the most expensive model); admins get alerts at 75% and 90% of spend limits. Anthropic blog

  • claude-code The steganography patch is confirmed shipped: Claude Code v2.1.197 (released July 1) removes the Unicode-swap tracking mechanism, though no mention of it appeared in the official changelog. The Register

Models & research

  • agentic-ai Zuckerberg told Meta employees at a July 2 town hall that AI agent development has not accelerated as expected over the last four months, and that the May restructuring (10% workforce cut, 7,000 reassigned to AI) was “not as clean” as planned. He expects meaningful returns within 3-6 months. US News/Reuters

  • agentic-ai Related context: Meta has now capped per-employee AI token spending after internal usage hit ~73.7 trillion tokens in a single month on an internal leaderboard called Claudeonomics. Employees are being steered toward MetaCode (proprietary) and away from external tools like Claude. MLQ AI

  • model-releases Together AI raised $800M Series C at an $8.3B post-money valuation (more than double its previous $3.3B). Round led by Aramco Ventures; Nvidia, Vista Equity, General Catalyst, Emergence Capital joined. Annual bookings crossed $1.15B last quarter on the back of tripling open-source model usage. TechCrunch

OSS

  • oss JuliusBrussee/caveman is the day’s biggest GitHub mover (2,851 stars). It’s a Claude Code skill — and works with 30+ other agents — that rewrites responses in terse “caveman-speak,” cutting output tokens by ~65% with six compression levels and companion commands for commits, reviews, and context compression. v1.9.1, 82.5k total stars. GitHub

  • oss obra/superpowers (1,205 stars today): agentic skills framework with a structured seven-phase software development workflow (design → plan → TDD → subagent execution → peer review → finalize). Works across Claude Code, Cursor, Copilot, Kimi Code. GitHub

  • oss safishamsi/graphify (937 stars today, v0.9.5 July 2, YC S26): converts any code folder into a queryable knowledge graph with interactive HTML, markdown report, and JSON output. Supports Claude Code, Cursor, Gemini CLI, Codex, 15+ agents. GitHub