AI INSIDER — ISSUE #43

April 24, 2026 | The practitioner’s edge on autonomous AI

The Hook

Two frontier model launches dropped simultaneously this week — GPT-5.5 from OpenAI and DeepSeek V4 from China — while a quiet infrastructure release may matter more to anyone actually running agents in production. Infisical shipped Agent Vault, an open-source credential proxy that changes how agents interact with secrets. Meanwhile, 75% of Google’s new code is now AI-generated, and a supply chain attack hit Bitwarden’s CLI. The gap between demo-grade and production-grade AI is closing fast, and the attack surface is expanding with it.

This Week’s Signal

Agents should never touch your secrets — Agent Vault makes that a guarantee.

Infisical launched Agent Vault this week: an open-source HTTP credential proxy that sits between your AI agents and every API they call. The architecture is simple but the implication is significant. Instead of retrieving credentials and handing them to an agent process, Agent Vault creates a scoped session and a local HTTPS proxy. The agent calls APIs normally — fetch("https://api.github.com/...") — and Agent Vault intercepts the request, injects the right credential at the network layer, then forwards it upstream. The agent never sees the secret.

Why this matters now: prompt injection is the attack vector everyone acknowledges but few have actually hardened against at the infrastructure layer. If your agent can be tricked into echoing its environment variables, you lose every API key it holds. Agent Vault eliminates that risk entirely — credentials are encrypted at rest with AES-256-GCM, keys are wrapped via Argon2id, and every proxied request is logged with method, host, path, and status (no bodies, no headers, no query strings).

It wraps any agent process: agent-vault run -- claude, agent-vault run -- codex, agent-vault run -- opencode. It supports Claude Code, Cursor, Codex, and anything else that speaks HTTP. Runs on macOS and Linux (x86_64 + ARM64). Single-line install: curl -fsSL https://get.agent-vault.dev | sh.

The HN thread (121 points on launch day) shows the practitioner community is watching. As agents take on more privileged operations — CI/CD triggers, database writes, billing API calls — brokered credential access isn’t a nice-to-have. It’s the same pattern the industry learned with human IAM a decade ago. Agent Vault applies it to non-deterministic systems that can be socially engineered.

3 Operator Playbooks

1. GPT-5.5 and DeepSeek V4 Hit Simultaneously — Choose Your Benchmark

OpenAI dropped GPT-5.5, its self-described “smartest and most intuitive” model yet (1,436 HN points, 953 comments). Hours earlier, DeepSeek released V4, claiming parity with GPT-5.5, Gemini, and Claude on major benchmarks (1,296 HN points, 927 comments). Two competing frontier model releases in one day is the new normal — and it puts every operator in a position of constant re-evaluation.

Your move: Do not trust vendor benchmark claims for your workload. Pull both models on your actual task distribution this week. The organizations winning with AI are running internal evals continuously, not waiting for third-party leaderboard updates. Treat GPT-5.5 and DeepSeek V4 as candidates to test, not conclusions to adopt.

2. Google’s 75% AI-Generated Code Number Is a Forcing Function

Sundar Pichai disclosed that 75% of Google’s new code is now AI-generated, up from 50% last fall. Anthropic reported a similar range internally: 70-90% via Claude Code. These are not aspirational numbers from pilot programs — they’re operational baselines at two of the most technically sophisticated engineering organizations in the world.

Your move: If your engineering team is not running a coding agent in daily workflows, you are now operating at a structural speed disadvantage relative to the frontier. This week: instrument one recurring development task — PR review, test generation, or boilerplate — through Claude Code or equivalent. Measure cycle time before and after. Ship the result, not the experiment.

3. Sullivan & Cromwell’s Three-Page Error List Is a Warning, Not a Punchline

Elite law firm Sullivan & Cromwell — counsel on the Trump cases and SpaceX/xAI merger — was forced to apologize to a federal judge after filing documents full of hallucinated case citations. The error list ran three pages long. This is not a story about lawyers being bad at AI. It is a story about deploying AI in high-stakes workflows without a verification step.

Your move: Any agentic workflow that produces artifacts consumed by humans in a high-trust context — legal filings, financial reports, compliance documentation, client deliverables — needs a structured verification gate. This means: (1) define what “correct” looks like before generation, (2) build a lightweight check that validates against that definition, (3) treat the agent output as a draft, not a deliverable. The cost of the check is trivial. The cost of the failure is not.

Steal This

Agent Credential Hardening Checklist — deploy before your next agent goes to production

PRE-DEPLOY AGENT SECURITY CHECKLIST
=====================================

[ ] Credentials never stored in environment variables visible to agent process
    -> Use Agent Vault or equivalent credential proxy
    -> Agent receives HTTPS_PROXY endpoint, not raw secrets

[ ] Prompt injection surface mapped and tested
    -> Does any user-controlled input reach a tool call that touches credentials?
    -> Test with adversarial inputs before production

[ ] Least-privilege API keys only
    -> Scope every key to minimum required permissions
    -> Rotate all keys used in agent sessions monthly

[ ] Request logging enabled
    -> Every outbound API call logged with method, host, path, status
    -> Retention policy defined (30 days recommended)

[ ] Output validation gate for high-trust artifacts
    -> Any agent output entering a human workflow reviewed before submission
    -> Especially: legal, financial, compliance, client-facing

[ ] Supply chain integrity verified
    -> Pin all CLI tool versions in lockfile (npm, pip, cargo)
    -> Checksum-verify agent runtime binaries on install
    -> Monitor for upstream compromise (ref: Bitwarden CLI / Checkmarx incident)

[ ] Session scoping enforced
    -> Agent sessions expire; no persistent long-lived privileged sessions
    -> Re-auth required for sensitive operation classes

Use this as a PR checklist for any new agent deployment. It takes ten minutes to review and has saved teams from the categories of failure that made headlines this week.

The Bottom Line

This week’s signal is straightforward: AI agents are moving fast into privileged territory — codebases, APIs, legal filings, enterprise inboxes — and the security infrastructure to support that transition is only now being built. Agent Vault is one answer to a problem that every production agent team has either already hit or is about to. GPT-5.5 and DeepSeek V4 landing the same day is a reminder that model selection is now a continuous operational decision, not a one-time choice. And Google’s 75% number should end any internal debate about whether AI-assisted development is real. The question is no longer whether. It is how fast, how safely, and with what verification layer standing between your agent and your most consequential outputs.

AI Insider is published by Digital Forge Studios Inc.