OpenAI and Meta both restructured their flagship AI products this week, Illinois signed the country’s toughest state AI safety law, a new prompt injection technique got a name, and Gartner put a hard number on how much of the agentic AI hype is likely to fizzle. Here’s what tech professionals need to know.
OpenAI Merges ChatGPT and Codex Into One App, Launches ChatGPT Work
On July 9, OpenAI folded its Codex desktop app into the main ChatGPT app on Windows and Mac, available on every plan including Free. Alongside the merger, OpenAI launched ChatGPT Work, a new agent built on GPT-5.6 that stays with a project for hours and hands back finished output, spreadsheets, slides, documents, or a working web app, instead of a chat transcript. It ships first to Pro, Enterprise, and Edu plans, with Plus and Business following in the next few days, and it includes a Compliance API for admins who need an audit trail of what the agent did.
Why it matters: The Compliance API is the detail worth paying attention to. Audit-trail gaps are one of the biggest reasons regulated teams stall out on agent adoption, so this is OpenAI addressing an enterprise blocker directly rather than just shipping a faster model.
Source: NxCode, ChatGPT Work and Codex Merge
Meta Launches Muse Spark 1.1, Starts Charging for AI for the First Time
Meta released Muse Spark 1.1 on July 9, an agentic coding and tool-use model with a 1 million token context window and support for parallel sub-agents that can operate desktop, mobile, and browser interfaces. Meta claims it topped rival models on the MCP Atlas and JobBench tool-use benchmarks. Notably, this is the first time Meta has charged businesses for API access to one of its models, pricing it at 1.25 dollars per million input tokens and 4.25 dollars per million output tokens. Meta also confirmed its custom “Iris” data center chip begins production in September.
Why it matters: Meta charging for API access at all is the bigger story here than the benchmark claims. It signals Meta sees agentic tooling, not just chat, as the product worth monetizing, which changes the competitive calculus for teams currently defaulting to OpenAI or Anthropic for agent workloads.
Source: TechCrunch, Meta Enters the Crowded AI Coding Battle
Illinois Signs the Nation’s Toughest State AI Safety Law
Governor JB Pritzker signed SB 315, the Artificial Intelligence Safety Measures Act, on July 6. It’s the first state law requiring annual third-party audits of frontier AI models and applies to systems trained with massive compute that generate over 500 million dollars in annual revenue. It takes effect January 1, 2027, and includes whistleblower protections for employees who raise AI safety concerns internally.
Why it matters: Combined with existing New York and California rules, Illinois’ law pushes roughly 40 percent of the US market under what’s effectively becoming a de facto national AI safety standard. If your company builds on or resells frontier models, this is worth a compliance review even if you’re not headquartered in Illinois.
Source: Capitol News Illinois, Pritzker Signs Landmark AI Regulation Bill
“Friendly Fire” Exploit Turns Coding Agents Against Their Own Security Checks
The AI Now Institute published a proof-of-concept this week called Friendly Fire, an attack against defensive coding-agent workflows that requires no hooks, plugins, or special configuration. It works by planting prompt injections in ordinary repository content, content the agent already trusts as part of its normal read access, and using that to persuade the agent to run malicious scripts during what looks like routine security testing.
Why it matters: This is a good reminder that prompt injection risk doesn’t require an exotic attack surface. If your coding agents read pull requests, issues, or README files as part of their normal workflow, that’s an untrusted input channel, whether your team has been treating it that way or not.
Source: NxCode, Friendly Fire and Rogue Agent
Gartner: Nearly Half of Agentic AI Projects Will Be Cancelled by 2027
A fresh Gartner analysis estimates up to 234 billion dollars in enterprise application software spend is exposed to disruption from agentic AI by 2030, roughly 20 percent of total enterprise SaaS spend. In the same report, Gartner forecasts that more than 40 percent of agentic AI projects launched today will be cancelled by the end of 2027, largely due to unclear ROI, governance gaps, and rising costs.
Why it matters: Both numbers can be true at once. Agentic AI is a real threat to legacy SaaS pricing models, and a lot of the agent projects being greenlit right now are going to fail anyway. If you’re pitching an agentic initiative internally, lead with a narrow, measurable pilot rather than a sweeping transformation story. It’s the pilots with a clear success metric that tend to survive the 2027 cull.
Source: Gartner Newsroom, $234 Billion in Enterprise Application Software Spend at Risk
What to Watch Next Week
Watch whether Plus and Business users actually get ChatGPT Work on the “next few days” timeline OpenAI promised, since staggered enterprise rollouts have slipped before. Also worth tracking: how many other states introduce an Illinois-style third-party audit requirement now that a concrete legislative template exists, and whether Meta’s move to paid API access for Muse Spark pressures Anthropic or OpenAI on pricing for comparable agentic tiers.
