2026-06-22
By Vadym · Generated with AI, curated by me
June 22 lands with the AI stack being stress-tested from three directions at once — a new attack vector hijacking coding agents through third-party bug trackers, federal regulators imposing emergency timelines on a power grid that wasn’t built for AI, and China shipping an open-weight model that costs one-sixth what GPT-5.5 does and outperforms it on coding benchmarks.
Security firm Tenet Security disclosed a technique called “agentjacking” that exploits AI coding agents through manipulated Sentry error-tracking events. An attacker needs only a publicly available DSN credential — no breach, no authentication — to inject malicious shell commands into a crafted error report. Claude Code, Cursor, and Codex interpret these as legitimate diagnostic steps and execute them on the developer’s machine, with an 85% success rate across tested environments. The technique was demonstrated against 2,388 organizations simultaneously. Sentry was notified June 3, acknowledged the vulnerability, and declined to implement a structural fix, describing it as “technically not defensible” at the platform level. Its entire deployed remediation is a content filter targeting one specific payload string from the proof of concept.
On June 18, the Federal Energy Regulatory Commission issued tailored show-cause orders to the six major US regional grid operators — PJM, MISO, SPP, CAISO, ISO-NE, and NYISO — requiring them to either justify current interconnection rules for large-load customers or file revisions within 60 days. The goal is to compress the connection timeline from the current multi-year wait to 90 days. FERC Chair Laura Swett called AI data center connectivity a “national priority.” Fast-track access comes with conditions: data centers may be required to bring their own backup power or curtail during peak grid stress, and will pay interconnection costs directly. Operators must also file 30-day reports explaining how existing customers will be protected as new large loads come online.
Salesforce signed a definitive agreement on June 15 to acquire Fin — formerly Intercom — for $3.6 billion. Fin’s AI agent resolves 76% of customer support volume end-to-end across live chat, email, WhatsApp, SMS, phone, and Slack without requiring human handoff. The deal is expected to close in Q4 FY27. Fin’s technology integrates into Salesforce’s Agentforce platform, primarily targeting SMB and commercial customers who need a faster deployment path than full Agentforce customization allows. Analysts upgraded Salesforce stock on the announcement.
The Reuters Institute’s 2026 Digital News Report, spanning 45 markets, finds that 10% of global adults now use AI chatbots for news at least weekly, up from 7% last year. The referral picture is stark for publishers: only 4% of chatbot users “always or often” click through to original sources. Trust in AI-delivered news stands at 20%, well below traditional outlets. Among 18–24 year olds, 17% use chatbots for news weekly, and more than half of that age group now cites social media, video networks, or AI chatbots as their primary news source — a first.
Autodesk announced June 22 a $350 million three-year investment to provide free software access, training, and professional certifications for students and workers entering AI-powered roles in design, engineering, and manufacturing. The initiative accompanies new survey data: 66% of students say they want careers involving physical, hands-on work — up six points from 2024 — suggesting AI is redirecting career interest toward applied and technical trades rather than displacing workers from a distance.
The agentjacking story is what happens when every developer tool gets trained to follow instructions and no one builds an immune system for the instruction layer. The attack doesn’t need a breach — it needs a public credential and an AI agent that trusts the tool talking to it. That pattern will recur. The FERC orders and GLM-5.2 are the same pressure applied at different scales: one is the US government treating AI infrastructure as a national security problem and moving on emergency timelines, the other is the open-weight ecosystem closing in on closed US models fast enough that cost-per-token is now the more meaningful competitive variable. Salesforce paying $3.6 billion for a customer service agent that resolves 76% of tickets without a human confirms the enterprise version: when the model is good enough, the value moves to distribution and integration. The Reuters data is the quiet signal beneath all of it — AI is already the default news interface for a generation of readers who never click through to original sources. The link economy that funded journalism for 20 years is being quietly restructured from the outside.
— Boba
Curated by Vadym