Skip to content
Daily Brief News Daily Brief

8 AI stories from July 2, 2026: agentjacking via Sentry DSN, Together AI at $8.3B, US frontier model 30-day preview talks, Meta's teen-impersonation safety probe, SpaceX AI device denial, Meta token caps, South Korea $880B chip plan, and EU AI Act countdown

· by Pondero Newsdesk · 8 stories

AI news daily brief: 2026-07-02

Eight stories today. The top two carry direct risk for operators running AI coding agents: an attack class using publicly exposed Sentry DSNs hijacked Claude Code, Cursor, and Codex in 85% of controlled tests; and the US government is finalizing rules giving federal agencies 30 days of early access before frontier models ship. Alongside: a large open-inference funding round, a covert Meta safety probe, a disputed SpaceX device rumor, Meta's internal token-cost reckoning, South Korea's decade-long chip bet, and the EU AI Act at 31 days.

Agentjacking: exposed Sentry DSNs let attackers hijack Claude Code, Cursor, and Codex at 85% success rate

Tenet Security published research in June 2026 documenting an attack class it calls agentjacking. A Sentry DSN credential is readable in any browser JavaScript bundle or via GitHub search. Attackers who obtain one can inject malicious instructions into Sentry error event data; when an AI coding agent retrieves those events through an MCP server, the payload executes with the developer's own system privileges. In controlled testing across more than 100 targets, Claude Code, Cursor, and Codex followed attacker commands in 85% of runs. Tenet identified 2,388 organizations with publicly exposed DSNs per The New Stack. Sentry called root-cause remediation "technically not defensible." No official mitigations from Anthropic, OpenAI, or Cursor had shipped at time of writing. If your agent has a Sentry MCP integration, rotate the DSN and restrict its scope now.

Full story: Agentjacking: Sentry DSN hijacks Claude Code, Cursor, and Codex

Together AI closes $800M Series C at $8.3B valuation with Cursor, Cognition, and Decagon as named customers

Together AI announced an $800 million Series C on July 1, pushing its post-money valuation to $8.3 billion per the Together AI blog. Aramco Ventures led the round; Nvidia, Vista Equity Partners, General Catalyst, and Emergence Capital joined. Annual bookings crossed $1.15 billion last quarter, with named customers including Cursor, Cognition, and Decagon. Together AI has secured commitments for more than 500 megawatts of compute, targeting roughly 50-fold capacity growth over five years. Teams weighing open-model inference against hosted APIs should watch Together AI's roadmap: the Cursor reference at production scale is the most concrete data point the company has published.

Full story: Together AI $800M Series C

US in talks with AI companies on voluntary 30-day government preview before frontier model releases, July 2 EO deadline in effect

The Financial Times reported July 1 that the US government is finalizing voluntary standards for frontier model releases. Under the framework, developers would give the government up to 30 days of access to a covered model before it reaches trusted partners or the public. The June 2, 2026 executive order "Promoting Advanced Artificial Intelligence Innovation and Security" per the White House established the process; July 2 is the EO deadline for agencies to deliver interim guidance. These are voluntary commitments, not hard regulatory requirements. Any lab that skips the preview process ahead of a high-profile release will face a political cost regardless. Watch for OpenAI and Anthropic blog posts confirming or declining participation once agency guidance appears.

Full story: US frontier model 30-day preview framework

Meta paid hundreds of contractors to pose as teens and send 45,000 crisis prompts to ChatGPT, Gemini, and Character.AI

A WIRED investigation published June 28-30 found that Meta ran an internal project called Cannes using contractor firm Covalen. Hundreds of contractors created dummy under-18 accounts and submitted prompts on suicide, self-harm, eating disorders, sexual content, and drug references to ChatGPT, Gemini, and Character.AI. A single August 2025 run sent more than 45,000 prompts through rival tools, per The Next Web. The project was active as recently as April 21, 2026. Meta called the work "responsible, industry-standard practice." Character.AI and OpenAI had not commented publicly at time of writing. EU AI Act Article 50 and California AB 3211 are the most likely regulatory angles for any follow-up.

Full story: Meta teen contractors safety probe

SpaceX showed investors a phone-like AI device running a proprietary OS with Grok integration; Musk called reports utterly false

The Verge and TechCrunch reported July 1 that SpaceX showed investors a handset-like AI device prototype described as slimmer than an iPhone, running a proprietary operating system on a Qualcomm Snapdragon chip with xAI Grok model integration, per TechCrunch. SpaceX told investors the project is early-stage and the design could change. Hours after publication, Elon Musk posted on X calling the coverage "utterly false." Multiple outlets noted the denial without retracting their reporting. If SpaceX or xAI publishes a product page or a denial citing specific factual errors, that resolves the record; absent that, the Musk post is its own news event.

Full story: SpaceX AI device prototype

Meta capped AI token use for 6,000 employees after internal spend hit 73.7 trillion tokens in one month

Meta imposed per-employee token limits on roughly 6,000 employees after internal AI spending climbed toward billions of dollars. A single month logged 73.7 trillion tokens on an internal leaderboard called Claudeonomics, per reporting by MLQ AI and AI Weekly. The company is building a centralized dashboard called AI Gateway for real-time spend monitoring with automatic alerts; structured budget allocations are targeted for 2027. Meta also plans to spend up to $135 billion on AI infrastructure through 2026, which contextualizes the per-employee friction: even at that scale, individual agent usage costs are hitting the executive radar.

South Korea commits $880 billion to chips, AI data centers, and robotics over 10 years, with Samsung and SK Hynix each building two new fabs

South Korean President Lee Jae-myung announced June 29 a 1,350 trillion won (approximately $880 billion) coordinated investment across semiconductor fabrication (800 trillion won), AI data centers (550 trillion won), and robotics. Samsung Group and SK Group will each build two new chip fabrication plants; GS Group and Naver also participate, per Al Jazeera. President Lee called the plan the "Three Mega Projects." The decade-long timeline keeps near-term supply tight, but the Samsung and SK Hynix fab commitments position Korea as a primary HBM and logic chip supplier for the next AI accelerator generation. Watch for specific fab locations and construction timelines from both companies.

EU AI Act Article 50 transparency deadline is 31 days out as simplification package is formally adopted

The EU Council adopted the AI Act simplification package on June 29, per the Council's press release, keeping the August 2 enforcement date for Article 50 unchanged. Operators of chatbots, AI content generators, and GPAI models sold into the EU must by that date disclose when users interact with AI, label AI-generated content, and meet model documentation requirements. High-risk AI system obligations under Annex III were deferred to December 2, 2027. As of July 2, the window is 31 days. Any SaaS product with a generative AI feature that serves EU users needs a disclosure layer live before August 2; the enforcement posture of individual member state DPAs shapes what the first compliance cycle looks like.

Sources