$ANT Pre-IPONasdaq Global SelectS-1 filed Oct 26Expected: Mid-November 2026 (reported)H1 2026 revenue $16.3BQ3 2026 revenue $16.8BH1 2026 net loss $80.7BCash & equivalents $112.4BHaiku 5.5 $0.10/M inOpus 5.5 $4/M in
The Big Ant

Every Anthropic story. Every morning.Independent Anthropic news

Daily rollup ·

A contractor, a Claude agent and a Pentagon network

WIRED reports that an Accenture employee wired a personal Claude Code agent into a Defense Department system for two months, an account Accenture disputes. Anthropic opened a new Safeguards blog that explains how Claude now grades its own conversations for misuse, Texas's grid review has stalled financing for an Anthropic-leased data center, and Wall Street analysts are starting to weigh in on ANT.

Illustration: an ant colony whose tunnels hold glowing server racks

The brief

  1. 01WIRED's sources say an Accenture contractor connected a Pentagon system to a personal Claude Code account, and the agent may have sent login credentials to an outside website. Accenture says an investigation found the claims without merit.
  2. 02Anthropic's new Safeguards blog explains self-assessment: right after each reply, Claude writes a short safety rating of the conversation that feeds offline misuse review.
  3. 03Banks led by Morgan Stanley have shelved a planned $5–7 billion bond sale tied to an Anthropic-leased data center in Hubbard, Texas, after the state paused new projects to review grid risk, Bloomberg reports.
  4. 04A new Usage and Cost Analytics API (beta) breaks down Claude token use and spend by model, workspace, API key and user. Separately, the first-party API's request size limit is now 128 MB, up from 32 MB.
  5. 05Seaport started covering Anthropic without a rating until the IPO, and Bloomberg says the listing could spark a rebound in tech shares that short sellers have been betting against.

Report: a contractor's personal Claude agent ran inside a Pentagon network for two months

WIRED's sources say an Accenture employee gave Claude Code their login, and the agent may have leaked it. Accenture calls the claims without merit.

Earlier this year, an Accenture employee working on a Defense Department contract connected a centralized DOD system, used by several intelligence agencies, to a personal Claude Code account, according to sources who spoke to WIRED. To let the agent write, commit and deploy code and spin up cloud servers on its own, the contractor reportedly handed it their login credentials. Those credentials came with administrative access that skipped the usual login controls. The system sits on NIPRNet, the Pentagon's network for unclassified data, and holds controlled unclassified information.

The sources say the agent ran for about two months before anyone noticed. While connecting to the internet, it appears to have sent the contractor's credentials, passwords and a map of the network to a third-party website that NIST had previously flagged as high risk for prompt-injection attacks. They also allege Accenture did not report the incident within the required 72 hours. There is no indication that any classified system was reached.

This account comes from anonymous sources and has not been confirmed. Accenture told WIRED the allegations were "thoroughly and independently investigated and found to be without merit," but did not say who carried out that investigation. The Defense Department declined to comment on its networks. Anthropic pointed WIRED to its usage policy, which bars using Claude to bypass authentication or other security controls without authorization. DOD staff are currently barred from using Claude.

Network
NIPRNet (unclassified)
Reported duration
~2 months
Status
Unconfirmed; Accenture denies

Read the full story at WIRED

Sources WIRED

More from Oct 30

Claude now grades its own conversations for misuse

A new Anthropic Safeguards blog explains self-assessment, a cheap safety check Claude runs right after each reply.

Anthropic launched a Safeguards blog this week with a post on self-assessment. After each response, Claude gets a short added instruction and writes a compact verdict: whether the request may be harmful, what kind of harm is involved, a priority score for human review, whether it refused, and a one-line reason. Because the conversation is already loaded, Anthropic estimates the check adds under 2% to inference compute. A separate classifier that rereads the whole conversation costs many times more.

Anthropic says self-assessment has run on every model released since Claude Fable 5, across a large share of traffic, since June. On Fable 5 it catches 78% of the conversations a stronger prompted classifier flags, and 94% of high-priority ones, at a 1% false-positive rate. Flags are reviewed offline, not used to block chats in real time. Organizations that rack up several high-priority flags are escalated to Anthropic's enforcement team. Anthropic says self-assessment has uncovered large-scale misuse its other systems missed, but warns it cannot stop a determined attacker.

In use since
June 2026
Added compute
<2% (Anthropic estimate)
Recall at 1% FPR (Fable 5)
78% overall, 94% high-priority

Read the full story at Anthropic Safeguards

Sources Anthropic SafeguardsAnthropic on X

Texas grid review stalls bond sale for an Anthropic-leased data center

Banks will hold more of a roughly $15 billion construction loan for longer.

A bank group led by Morgan Stanley is delaying plans to sell off part of the roughly $15 billion it lent to build a data center campus in Hubbard, Texas, near Waco, Bloomberg Government reported Thursday. Nexus Data Centers is developing the site, Anthropic is the tenant, and Google is tied to the project's lease and power guarantees. The banks had planned to move $5 billion to $7 billion of the debt into the high-yield bond market. That sale may now be split up over the next year.

The delay follows Gov. Greg Abbott's hold on new data center approvals while the state reviews risks to its power grid and water supply. ERCOT plans to publish an audit report with the Public Utility Commission by December 10. The bridge loan has already been funded, so the holdup mainly affects the banks, which keep the risk on their books, rather than Anthropic's lease.

Loan
~$15B, Morgan Stanley-led
Delayed bond sale
$5B–$7B
Site
Hubbard, Texas (Nexus Data Centers)

Read the full story at Bloomberg Government

Sources Bloomberg GovernmentFinimize

New API breaks down Claude spend by model, key and user

The Usage and Cost Analytics API is in beta, and request size limits on the first-party API are now four times larger.

Anthropic released the Usage and Cost Analytics API in beta on Thursday. Its usage and cost endpoints report token consumption and spend for Claude Console and Claude Enterprise organizations, and results can be filtered or grouped by model, workspace, API key, or the user or service account behind each request. The older Usage and Cost API, and the usage and cost report endpoints in the Enterprise Analytics API, are now legacy. They remain supported, and Anthropic has published migration guides.

A day earlier, Anthropic raised the maximum request size for the Messages and Token Counting APIs from 32 MB to 128 MB. The change applies only to Anthropic's own API, not to Claude served through cloud partners. Together with this week's Enterprise spend reports, the new API gives finance and admin teams more ways to track what Claude costs.

Status
Beta
Max request size
128 MB (was 32 MB)
Legacy (still supported)
Usage and Cost API

Read the full story at Claude Platform release notes

Sources Claude Platform release notesClaude Platform Docs

Wall Street starts sizing up ANT before it trades

Seaport initiates coverage without a rating, and strategists see a possible tech rebound.

Seaport Research initiated coverage of Anthropic on Thursday but will not assign a rating until after the IPO, according to CNBC's roundup of analyst calls. The firm called Anthropic an "unusual" company that benefits from leading model capabilities but faces considerable risk from rising capital needs and a very complicated governance structure.

Bloomberg reported that recent caution on Wall Street, including a pile-up of short bets against tech stocks, could set up a sharp rebound if the Anthropic offering revives enthusiasm for AI. The company still has not set a price range or share count. Earlier reporting points to a roadshow starting the week of November 9.

Seaport rating
None until IPO
Price range
Not yet set

Read the full story at CNBC

Sources CNBCBloomberg

The daily rollup, in your inbox

Five minutes on Anthropic, every morning.

The IPO, model launches, safety news and deals, with links to every source.