Access Denied

I round up the most relevant AI-in-finance news - the deals being done, who’s rolling out what, and what’s actually working on the front lines

…It's called Mythos, and within 72 hours the US Treasury Secretary and the Fed Chair had summoned Wall Street bank CEOs to Washington for an emergency briefing. In the same week, news broke of a $200 million joint venture with Blackstone, General Atlantic and Hellman & Friedman to embed Claude across PE portfolios.

Elsewhere: Apollo called for a full reset in how software gets underwritten, SAP scrapped seat-based pricing, and Goldman said AI has been cutting sixteen thousand US jobs a month.

But first, my take on what Mythos actually tells us about the next few quarters.

In This Week’s Issue:

From The Trenches

Access Denied

Mythos dropped on April 7. In testing, it discovered thousands of previously unknown security flaws across every major operating system and web browser, including one in OpenBSD that had gone undetected for 27 years. To put that in context: OpenBSD is considered one of the most security-hardened systems in existence, and this flaw had survived decades of expert human review and millions of automated security tests. Mythos found it without supervision.

Within 72 hours Anthropic had locked the model behind a curated list of eleven vetted organisations, Scott Bessent and Jerome Powell had summoned Wall Street CEOs to Washington for an emergency briefing, the Bank of England had added it to its Cross Market Operational Resilience Group agenda, and the Bank of Canada had met with banks and financial firms in Ottawa.

Forget the model for a second. Think about the weight of that response. Treasury secretaries don't summon bank CEOs for a software release. Central banks don't convene cross-market resilience groups over a product launch. This is AI capability being treated, for the first time, with the same institutional seriousness as a systemic financial risk. That alone tells you where we are.

The immediate story is security. Mythos found vulnerabilities that had been hiding in plain sight for decades, and Anthropic decided the risk of public release outweighed the benefit. Fair enough.

But the more interesting story is what comes after. Mythos is the first model a frontier lab has publicly decided is too capable to ship. That sets a precedent. It means there is now a confirmed gap between what exists at the frontier and what gets released to the public API.

That gap is only going to widen. The security rationale applies today. But capability gating doesn't stay limited to security for long. Once the infrastructure exists to restrict access, it gets used for commercial reasons too. Tiered access. Preferred partners. Strategic relationships. The playbook writes itself.

Now look at the rest of the field. Meta launched Muse Spark this week. Sub-Opus on raw capability, but genuinely competitive on the benchmarks that matter for day-to-day enterprise work. Open-weight models are running comfortably on consumer hardware and even on phones. As I wrote last week, DeepSeek V4 is a trillion-parameter architecture that will run entirely on Huawei Ascend chips, no Nvidia at all.

The public tier is getting very good. For most of the practical workloads a finance team cares about, model capability is no longer the limiting factor. You can do serious work with what's freely available right now.

But what Mythos reveals is that there's a layer above that public tier, and access to it is going to be relationship-gated. We were always heading in this direction, but Mythos accelerated the timeline. The divide between what's publicly available and what's available to vetted institutions is going to deepen faster than most people expected.

The same week Mythos dropped, Anthropic launched Managed Agents, a hosted service that takes care of all the infrastructure involved in deploying agents: the sandboxes, the orchestration, the context management, the tool routing. You bring the use case, they handle everything else. It's Anthropic making it as frictionless as possible to deploy agents, but only if you're building on Claude.

Read that alongside the $200m PE joint venture and the Mythos gating, and a pattern emerges. All three moves are about consolidation. Anthropic isn't just building better models. They're locking in distribution, reducing the friction to adopt Claude specifically, and creating barriers to switching away. It's further proof that the models themselves are good enough now. The race has shifted to who captures the deployment layer.

And that's the takeaway for deal professionals. The models are good enough. They have been for a while. The bottleneck is everything else around them: the implementation, the workflows, the integrations, the people who know how to make it all work inside your firm. The firms still waiting for models to get better before they commit are losing ground to the firms that started building six months ago. The gap isn't in the model anymore. It's in who's figured out how to use it.

News Digest

Why Officials Are So Worried About Mythos

Bloomberg published a detailed explainer this week on why Mythos has governments and regulators this concerned. The short version: Mythos can autonomously discover and exploit security flaws in software at a scale and speed that no human team can match. It found thousands of "zero-day" vulnerabilities, flaws that software developers didn't know existed, across every major operating system and browser. In one test, a non-expert asked Mythos to find a way to remotely take control of a computer, left it running overnight, and came back the next morning to a complete, working exploit.

Anthropic has restricted access to eleven vetted organisations under a programme called Project Glasswing, including Amazon, Apple, Google, Microsoft, Nvidia, CrowdStrike, Palo Alto Networks, JPMorgan and the Linux Foundation. The idea is to let these organisations use Mythos defensively, finding and patching their own vulnerabilities before attackers can exploit them.

The details:

Why it matters: This is the first time a frontier AI lab has publicly withheld a model on capability grounds. It sets a precedent for how the most powerful AI gets distributed going forward, and it puts cybersecurity squarely on the agenda for every board and IC in finance.

My take: The defensive potential is real. But so is the asymmetry problem. Less than 1% of what Mythos has found has been patched. Meanwhile, Palo Alto's CEO warned that within six months, a single attacker will be able to run campaigns that previously required entire teams. The transitional period, where AI-powered offence moves faster than AI-powered defence, is the dangerous bit. For PE firms, the immediate question is whether your portfolio companies' cybersecurity posture was underwritten for a world where AI can find and exploit decades-old flaws overnight. For most, the answer is no.

Read more

Anthropic's $200m PE Joint Venture

The Wall Street Journal reported this week that Anthropic is in talks to commit around $200 million to a joint venture backed by Blackstone, General Atlantic and Hellman & Friedman. The vehicle would act as a dedicated consulting and implementation arm, deploying Claude-based tools directly inside PE-owned portfolio companies. Early reporting suggests it could eventually scale to a $1 billion fund, with Anthropic supplying the models and the PE firms supplying capital, distribution and relationships.

The details:

Why it matters: Anthropic has decided the bottleneck for enterprise AI adoption is implementation, not model capability. So rather than just selling API access, they're building a services business to do the hard work of making it stick inside real companies.

My take: A frontier AI lab putting $200m into a consulting business tells you something about where the value has shifted. The model works. Getting it deployed inside messy, real-world organisations is the constraint. That's the same conclusion Goldman reached when they embedded Anthropic engineers for six months. The question for PE firms is whether you want Anthropic as your implementation partner or your competitor. Because right now, they're positioning to be both: the model provider and the consultancy that deploys it. Firms that build their own capability to integrate and orchestrate AI, independent of any single model provider, will have more flexibility as this market evolves. The ones that outsource the whole thing to Anthropic's JV are buying convenience today at the cost of optionality tomorrow.

Read more

Other Interesting Things I’ve Read of Seen This Week:

CoreWeave lands $21B Meta deal and multi-year Anthropic agreement back-to-back (April 9-10) - Meta expanded its contract to $35.2B total through 2032. Anthropic signed on the following day. Revenue backlog hit $66.8B. Stock surged 24% in a week. The AI infrastructure capital story now has its own gravitational pull.

OpenAI halts £31bn Stargate UK project (April 9) - Energy prices and regulatory deadlock killed the UK leg of Stargate before it broke ground. Turns out "we have infinite money" still runs into "you can't build a substation in Oxfordshire."

Canva acquires Simtheory and Ortto (April 8) - Canva bought an agent management platform and a marketing automation company on the same day. When a design company buys two martech platforms in one announcement, the "we're a design company" era is officially over.

Gen Z workers are intentionally sabotaging their company's AI rollouts (April 8) - A Writer and Workplace Intelligence survey found a meaningful share of employees admit to quietly breaking internal AI deployments. The human rate limit on AI adoption is more literal than people think.

Apollo calls for a full reset in software underwriting (April 7) - Apollo's research team argues AI disruption and higher rates require a rewrite of how software gets valued. Same week SAP moved to consumption pricing and Marathon warned of a default wave in PE-backed software credit. When Apollo, SAP and a credit fund all land on the same point simultaneously, it's not a coincidence.

Anthropic acquires Coefficient Bio for ~$400m (April 7) - Anthropic's first move into biology, picking up a stealth AI-for-biotech startup in an all-stock deal. The Frontier Red Team now sits next door to the people doing protein design. Busy week for Dario.

AI is cutting 16,000 US jobs a month, Gen Z taking the brunt (April 6) - Goldman's economics team put net US job losses attributable to AI at around 16k per month, concentrated in entry-level knowledge work. The analyst pyramid is being eroded from the base up.

Acquisition Intelligence is a weekly newsletter on AI in M&A for finance professionals, private equity investors, investment bankers, corp dev teams, and deal-makers.

For questions, feedback, or to share what you're seeing in the market, reply to this email.

P.S. I'm Harry, co-founder of DealSage. We're building an AI-native deal intelligence platform to help professionals turn their institutional knowledge into better decisions. If you're curious what we're up to, check out dealsage.io or just reply here