This briefing is informational only and does not constitute legal or financial advice. It reports on public industry developments and does not reference any Tienta client.
This Week in AI
Platform Roundup
- Anthropic (Claude): On October 26, Anthropic announced it had filed a registration statement on Form S-1 with the SEC for a proposed initial public offering of its Class A common stock. The next day it announced an expanded agreement with the U.S. General Services Administration that gives government partners a 40 percent discount on Claude Enterprise, Claude for Government, and the Claude API.
- OpenAI (ChatGPT): GPT-6.1 Terra launched October 26 for Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex, and in the API at $1 per million input tokens and $5 per million output. OpenAI describes it as built for fast, precise computer use, such as filling out forms and working across apps. It says Terra completes tasks over three times faster than GPT-6 Astra and GPT-6.1 Sol, and that its error rate on an internal benchmark at maximum effort is 4.3 percent. Those are OpenAI's own figures, and the company still recommends Sol or Astra for the most complex coding.
- Microsoft (Copilot): On October 28, Microsoft introduced the Microsoft Copilot harness, the agent framework that GitHub's engineers built, as a managed service in Microsoft Foundry. It supports more than 20 frontier models plus your own API key. Microsoft runs the agent loop and its sandbox, and you supply the instructions, tools, and choice of model. The announcement gave no pricing.
Beyond the Platforms
Google released its newest model to cyber defenders first and has not said when everyone else gets it
Gemini 4 Carbon, announced October 28, is described by Google as its strongest model, with state-of-the-art cybersecurity and long-horizon capabilities. Access begins with a subset of cyber defenders and expands gradually to trusted partners through Google's Fairwind Program. Developers, businesses, and consumers follow "as soon as possible," with no date given. Google says it is taking part in the U.S. government's voluntary pre-release model testing. Pricing will match Gemini 4 Argon's introductory $2 per million input tokens and $10 per million output, and a footnote says rates rise to $4 and $20 after the introductory period. The benchmark results Google published, including 88.2 percent on a vulnerability-discovery test, are its own and were evaluated without safety guardrails.
Why it matters: Google is giving defenders first access to a model built to find software flaws. For most businesses the practical effect is a wait. The strength cuts both ways, because the same ability that finds bugs for a defender can find them for an attacker, and a staged release is how a lab limits that risk. If tools like this do find more flaws, the visible result for you would be more patches and vendor advisories to keep up with.
Industry Spotlight
For Legal Practices
New York's courts told lawyers to verify AI work and document it, with no disclosure required
On October 27, New York's court system released "The Risks of Using Artificial Intelligence in Preparing Court Papers: What Every Attorney Should Know," a seven-page guide from its Advisory Committee on Artificial Intelligence and the Courts. It urges attorneys to confirm all AI-generated citations and quotations through authoritative legal research tools and to document the steps they took to verify them. Attorneys are not required to disclose AI use when filing or to file a separate certification beyond the implicit certification made when signing. Chief Administrative Judge Joseph Zayas said attorneys "remain fully accountable for the accuracy of the papers they submit to the courts." The guide leaves existing professional conduct rules and the state's frivolous-conduct sanctions rule in place. The coverage we reviewed does not say whether the guide is binding.
Why it matters: New York and California are now taking different approaches. California's SB 574, effective January 1, 2027, requires disclosure to the court when generative AI was used. New York requires neither disclosure nor certification, but expects verification you can show. A firm that practices in both places is better off with one standard that satisfies the stricter rule: verify every authority, keep a record of who did it, and be ready to disclose where a court requires it.
For Financial & CFO Firms
The AICPA proposed a rewrite of the SOC 2 criteria, and comments are open until January 15
On October 27, the AICPA released an exposure draft of the 2027 SOC 2 Trust Services Criteria for security, availability, processing integrity, confidentiality, and privacy. According to PwC's summary, the revisions are meant to improve the criteria's relevance, clarity, and usability in response to evolving technology and third-party risk. This is the first of three phases, and comments are due January 15, 2027. The draft is not specific to AI. We include it because SOC 2 is the report firms ask software vendors for, and the AICPA's pilot with the Artificial Intelligence Underwriting Company, covered in issue 11, builds an AI certification on top of it.
Why it matters: if your firm relies on vendors' SOC 2 reports, or issues them, the criteria that define what a clean report means are about to change. The comment window is open until January 15. At a minimum, ask your key AI and software vendors whether they have a current SOC 2 report and what it covers.
For SaaS & Engineering Teams
The built-in browsers in five AI coding tools can be turned against the developer
Help Net Security reported October 29 on research by Forever Security, which calls the pattern EMBRACE, an embedded browser trusted like a real one. The researchers tested Cursor, GitHub Copilot, OpenAI's Codex, Microsoft's VS Code, and Anthropic's Claude Desktop, and in all five a basic browser safeguard gave way. In Codex, VS Code, and Claude Desktop, a project disguised as a dull report led the assistant to offer a link to a local HTML file, and one click let a background script read files across the machine. On the researchers' own machine that collected SSH keys, environment files, and npm tokens. In Cursor, a web page's hidden command led the agent to turn its own sandbox off and retry after the sandbox blocked it. GitHub fixed Copilot in version 1.1.29 on October 13, and Cursor patched on October 1. As of October 29, Codex and VS Code were unfixed. Microsoft declined the report, citing Workspace Trust, and OpenAI had not replied to a July report. Anthropic first declined Forever's report, then accepted a similar one from Noma Security, and the article does not say whether Claude Desktop is fixed.
Why it matters: last week's story was a stray instruction file in Downloads, and this one is a stray HTML file in a project. The common thread is that anything an AI coding tool opens, reads, or links to can carry instructions or code, and the person clicking is trusting the assistant's suggestion. Treat projects you did not write as untrusted, keep these tools updated, and do not assume a sandbox will hold when an agent is the one deciding whether to use it.
On the Horizon
Google's introductory pricing for its Gemini 4 models has a built-in end. Gemini 4 Carbon will launch at the same introductory rate as Gemini 4 Argon, $2 per million input tokens and $10 per million output, and Google's footnote says $4 and $20 apply after the introductory period. Google's post does not say when that period ends, and Carbon is not yet open to most developers.
Why get ahead of this now: if you build a product or workflow on an introductory price, work out the cost at the higher rate before you commit, so a price step does not become a surprise.
Try This This Week
List every AI coding tool on your team's machines, including Cursor, GitHub Copilot, Codex, VS Code, and Claude Desktop, with its current version. Update the ones with fixes available, starting with Copilot 1.1.29 or later and a current Cursor build. For Codex and VS Code, have each developer review which projects they have marked as trusted, and remove any they do not recognize.