Laziness Is Destroying the Software Industry

Here is what happened to the software industry in the last seven days: npm leakage. Mercor leak. LiteLLM compromised. Claude Code's entire source code leaked. Railway CDN problem. Axios package compromised. Delve leaked their customer data.

Seven major security incidents. Seven days. Not in some obscure corner of the internet. In the core infrastructure that millions of developers and businesses depend on.

The Claude Code leak happened this morning. Anthropic, one of the most well-funded and technically sophisticated AI companies on the planet, accidentally published over 500,000 lines of proprietary source code to a public registry. Internal feature flags. Unreleased product roadmaps. System prompts. A source map file that should never have left a build server was sitting in plain sight on npm because someone's toolchain didn't strip debug artifacts before publishing. During the same window, the axios package was compromised with a remote access trojan. And last week, LiteLLM, a Python library with 3.4 million downloads per day, was backdoored by threat actors who stole credentials through a prior attack on Trivy, a security scanner. The attackers compromised a security tool to compromise an AI tool. The malicious versions harvested SSH keys, cloud tokens, Kubernetes secrets, and anything else they could reach.

None of this should be normal. All of it is becoming routine.

Agentic AI is failing in production. We see it every week.

We spend our days in conversation with operations leaders at insurance carriers, financial institutions, logistics operators, and healthcare companies. We just returned from the Gartner Data and Analytics Summit, where we spoke with over 200 enterprise teams in a few days. The consistent pattern: companies that deployed agentic AI systems into critical workflows are pulling them back. Not because the demos were bad, but because the production outcomes were unacceptable.

Gartner predicts that 40% of agentic AI projects will be canceled by 2027. We think that's conservative. In the industries we serve, an error rate that would be considered negligible in a chatbot is catastrophic in a claims processing pipeline. A 0.5% error rate on insurance claims at scale means millions of dollars in liability. Not a rounding error. A business-ending event.

At the Gartner Summit, an executive at a major logistics company asked us point-blank: "Are you guys an agent?" We said no. And he said, "Thank God." That is the first time in the history of this company that someone was relieved we weren't an agent. The market is telling us something.

The companies that moved fastest to deploy agents into critical operations are now the ones scrambling to explain to their boards what went wrong. The error rates are too high. The audit trails don't exist. The systems make the same mistakes over and over because there's no feedback loop, no operator oversight, no mechanism for the system to learn from corrections. They optimized for speed and got chaos.

The speed-versus-discipline debate is a false binary. Speed without verification is just recklessness.

The industry's current belief system goes like this: move fast, ship AI-generated code, iterate in production. The companies that move fastest win. Everyone else gets disrupted.

We think this is exactly backwards. The companies that will dominate the next decade are the ones that learned how to move fast and verify every step. Speed without discipline is not a competitive advantage. It is a liability that hasn't materialized yet. The Moltbook breach, where 1.5 million API keys were exposed because an AI-scaffolded database was deployed to production without review, is what happens at the end of that road. The LiteLLM supply chain attack, where nobody audited the dependency chain, is what happens at the end of that road. The seven incidents in seven days are what happens when an entire industry is on that road simultaneously.

The fastest typist in the room is not the best writer. The fastest coder is not the best engineer. And the fastest-deployed AI system is not the most valuable one. The most valuable one is the one that works correctly every single time, and gets better every time it doesn't.

Here's what enterprise AI actually looks like in 12 months.

By March 2027, the distinction between "AI companies that verify" and "AI companies that don't" will be the primary purchasing criterion for enterprise buyers. Not model quality. Not speed. Not price. Verification.

This is already happening. Every enterprise conversation we have ends with the same question: can I audit this? Can I trace every decision? If the system gets something wrong, does it learn from it, or does it make the same mistake a thousand more times before anyone notices? The companies that can answer yes are winning deals. The companies that can't are getting disqualified in security reviews before they ever reach a technical evaluation.

Within 12 months, the market will have a name for the distinction. We believe that name is Verified AI: AI systems where every decision is auditable, every output is traceable to its source, every error is caught by a human operator and permanently eliminated, and every step in the pipeline is constrained and observable. Not an agent that tries things and hopes for the best. A controlled, deterministic system that earns trust through transparency and gets better with use.

Verified AI is not a feature. It is an architecture. It requires building from the ground up with the assumption that every AI decision must be provable. That is a fundamentally different design philosophy from the one that dominates the market today, which is: call a model, hope the output is right, ship it.

The companies that deployed without auditability are about to learn the most expensive lesson in enterprise software.

Here is the consequence that nobody in the industry is saying out loud: the companies that deployed agentic AI systems into regulated workflows without built-in auditability are going to face regulatory action. It is not a question of if. It is a question of when.

When an AI system makes a decision that affects a consumer, whether that's approving a claim, triaging a medical authorization, or underwriting a loan, the organization is liable for that decision. If the system cannot explain how it arrived at the decision, if there is no trace, no audit trail, no mechanism for a regulator to inspect the reasoning, the organization is exposed. And regulators are paying attention. State-level AI legislation is accelerating. The EU AI Act is in enforcement. Industry-specific regulators in financial services and healthcare are issuing guidance on AI governance that explicitly requires explainability and auditability for automated decisions.

The companies that built on agentic systems, where the AI explores and experiments its way to an answer, are going to discover that their architecture is fundamentally incompatible with the regulatory environment that is taking shape right now. Retrofitting auditability into an agentic system is not a product update. It is a rebuild. And by the time the enforcement actions start arriving, it will be too late to rebuild.

What the industry is not saying publicly

Here is what we hear in private that nobody is saying on stage at conferences.

Enterprise teams are not worried about AI being too slow. They are worried about AI being ungovernable. The fear is not that the technology doesn't work. The fear is that it works well enough to deploy but not well enough to trust, and that the gap between "deployed" and "trusted" is where the career-ending mistakes live.

We hear this from CTOs at Fortune 500 companies who are under enormous pressure to "do something with AI" and are terrified of being the person who approved the system that caused the breach or the compliance violation. We hear it from operations leaders who watched their teams deploy AI-generated workflows that nobody on the team fully understands. We hear it from legal and compliance teams who are asking their technology counterparts: "If this system makes a wrong decision, can you explain to a regulator exactly why it made that decision?" And the answer, overwhelmingly, is no.

The honest reality of enterprise AI adoption in 2026 is that most deployments are not in production. They are in pilot. And most pilots are not graduating to production because the trust infrastructure doesn't exist. The gap is not in model capability. The gap is in verification, auditability, and the ability to prove that the system does what it claims to do.

We built bem because we thought the industry would arrive here. It has.

We are not writing this because we enjoy being contrarian. We are writing this because we have spent the last two years building the infrastructure that solves the problem everyone is now waking up to.

Every AI decision at bem is guardrailed, observable, and traceable. We don't use prompts. We use functions: constrained, auditable primitives that each do one specific thing. You can chain them together to build complex workflows, but every step is so controlled that if the system makes a mistake, it is immediately visible. An operator flags it, and that mistake is permanently eliminated. The system trains itself from corrections. Your team manages by exception rather than by the rule.

We are SOC 2, HIPAA, and GDPR compliant. We run annual third-party penetration tests. Code changes require peer review, automated testing, and static analysis before production. We encrypt everything. We don't vibe code our product. We don't cut corners because AI made it easy to cut corners.

We are the boring choice. We are the bricks. We are the production infrastructure that enterprises trust with their most sensitive operations because we do the work that doesn't make for good demos but makes for good outcomes.

Seven incidents in seven days. And next week there will be more.

The question for every company building on AI right now is not how fast you can ship. It's whether what you shipped can survive scrutiny. From a customer. From a regulator. From the next supply chain attack.

We built bem to survive that scrutiny. That's the whole point.

Antonio Bustamante is Co-founder and CEO of bem. Upal Saha is Co-founder and CTO. bem is the production infrastructure layer for unstructured data, trusted by enterprises in insurance, finance, logistics, and healthcare. SOC 2, HIPAA, and GDPR compliant.