Bullpen.News From Edwin Dorsey
Get Daily Recap Feedback

Wednesday, November 4, 2026

Back to the front page

From X

@convequity 1 click

Convequity on X

Anthropic and Binance show why the speed of containment matters. Cathedral, an AI security startup, says its agents found routes into Anthropic's internal messages and employee records by exploiting third-party business software during a security test. Anthropic acknowledged limited sales, HR and IT exposure, but says its model weights and code were safe, its monitoring caught some activity, and access was cut the same day. I lean toward Cathedral on the significance of that exposure. Internal information can help attackers impersonate colleagues and target people with privileged access. The damage can extend beyond the data initially exposed. Supply-chain risk becomes harder to contain as more tools and agents gain access to company systems. Tighter permissions should help. Building more tools internally might reduce vendor exposure, but it also transfers maintenance and security work to your own team. Binance's October 29 hack involved a different suspected entry point. Its preliminary investigation says a compromised DevOps engineer's laptop may have let an attacker reach an internal service and submit unauthorized withdrawals. Binance estimates a $6.3bn impact, with roughly $3bn frozen so far. An attacker using a trusted employee's access can cause damage without any malicious insider. And TRM Labs says most of the funds leaving Binance moved in about 20 minutes. That is why I think shift-right security becomes increasingly important: detecting and stopping attacks while systems are running. Security teams need to recognize lateral movement, where an attacker moves between accounts or machines, and contain it before the attacker reaches systems that can cause much greater damage. This is part of why we like Palo Alto Networks ($PANW). Cortex combines network, endpoint and cloud data with response controls, including the ability to isolate compromised devices. Connecting an investigation to an action that stops the attack matters when the attacker can move this quickly. It also makes cybersecurity an interesting market for premium-speed AI inference. An AI security agent needs to investigate accurately and act quickly, with current data and permission to intervene. Where the model's reasoning time is delaying containment, paying even 100x more per inference call could make economic sense if the reduction in expected losses exceeds the extra cost. Faster inference becomes valuable when it lets the defender stop an attack sooner without sacrificing accuracy.
@convequity

Convequity on X

Spirit's data auction shows what office AI is missing. Google ($GOOGL) opened at $5m and was selected at $10m for Spirit Airlines' internal data in August. Mercor was the $7.5m backup. Then micro1 proposed $12.5m. The sale still awaits court approval, with a hearing scheduled for November 6. The material includes emails, Teams conversations and business records. Google's counsel confirmed that it wants the data for AI training. I think this competition for records of real work will keep pushing up the value of enterprise data. This fits our earlier argument about why coding agents advanced faster than office automation. Coding has public repositories and tests that provide fast feedback. In office work, the context behind a decision is often scattered across private documents and conversations. Knowing how to use a spreadsheet is only part of the job. Knowing why an exception was approved, and what happened afterward, is much harder to learn from the public internet. My read is that Google is preparing for that gap. Internal records could help models learn how work moves through an organization, provided those records can be turned into tasks with outcomes that can be checked. Also it remains an open opportunity for $MNDY https://t.co/Rmzyg6DE0y, and other AI-native productivity startups to tacle. Gemini 4 Carbon already leads Artificial Analysis' AutomationBench-AA, which tests simulated workflows across business applications. Google is competitive in the kind of work this data could help models learn. Anthropic's IPO filing also puts adoption in perspective. As of September, over 50% of the Fortune 500 generated at least $100,000 each in trailing-year revenue for Anthropic. Among the publicly listed customers in that group, spending on Anthropic averaged under 0.1% of total expenses in 3Q26. We still think enterprise AI is falling short on office automation. Having access to a model tells us little about how much work a company can delegate to it. I find Google's enterprise opportunity more compelling over time than Meta's emphasis on personal AI. Google already has Workspace and Cloud as routes to customers. If better training on real work makes its models reliable enough to complete office tasks, it has a way to sell that capability. The unlock still requires models that can carry a task through the office environment with limited supervision. That remains some distance away. I think acquiring real work histories now could be a fruitful long-term bet for Google.

Investor Letters

All letters

New Ideas From The Road

Aurelion Research increased healthcare exposure and sees opportunities in banking and defense, alongside research on Alcoa, Nike, ONON and…

AA $10BNKE $60B
Aurelion Research 4 clicks

Oracle: The Party Is Almost Over

Oracle’s CDS spreads exceed 320 basis points as debt approaches $135 billion and fiscal 2027 capex targets $95 billion, amid heavy reliance…

ORCL $363B
David Desjardins 1 click

Site Centers (NYSE:SITC)

Site Centers' $3 per-share return-of-capital distribution and December delisting leave a discounted liquidation stub, with two properties…

CURB $3.6B
Real Assets, Real Value 4 clicks

Coupang: 3Q26 Business Recap

Coupang’s product-commerce growth rose 9% as data-leak effects fade, while management expects margins to return to pre-incident levels by…

CPNG $27B
Speedwell Memos 3 clicks

update on Gartner

Gartner’s growth has stalled after federal contract losses and cautious commercial spending, raising questions about its competitive…

IT $13B
scuttleblurb 1 click

Recycle Bin

Used aircraft engines can be repurposed as gas turbines, easing data center power constraints while tightening aviation supply.

EQT $33BRRC $12B
Doomberg 8 clicks
AMRZ · long

Amrize

Amrize (AMRZ) produces aggregates, cement and roofing materials across the US and Canada.

AMRZ $22B
Value Investors Club 3 clicks

TWLO Q3 FY12/26 Earnings Review

Twilio's Q3 revenue rose 20% to $1.56 billion, but Q4 organic growth guidance fell to 11% to 12% and RPO declined 28%.

TWLO $33B
Cestrian Capital Research 3 clicks

Stevanato down 30%. Yikes.

Stevanato fell 30% after forecasting low-single-digit 2027 growth as lower Nexa syringe volumes from a GLP-1 customer offset solid Q3…

STVN $3.6B
Best Anchor Stocks 3 clicks

Publications we follow

Every publication on this site, hand-picked.

129 publications

Newsletters 109

YouTube 4

Blogs 15

Seeking Alpha 74 of 276

Nominate a publication

X accounts we follow

Every account on this site, hand-picked.

91 accounts
Nominate an account

Investor letters we follow

359 firms

Sectors

Ordered by how often our editor reaches for them.