The reading order
-
1.
SOC 2 Is a Revenue Tool, Not a Security Tool
$25–45K total over 90 days. Scope it as a sales project, not a security project. The playbook for unlocking enterprise pipeline before Series A.
-
2.
php2shell: What to Do if Your Servers Run PHP
Two PHP runtime flaws, CVE-2026-109618 and CVE-2026-109619, both CVSS v4.0 9.2, fixed on October 21, 2026 in 8.5.12, 8.4.27, 8.3.36 and 8.2.35. The triage order: inventory every host that answers HTTP through PHP (php-fpm on localhost behind nginx counts), patch and restart the workers, apply php.net's partial mitigations, then isolate or go offline as CISA, the NCSC and the NSA describe. About 48.5% of PHP websites run a branch with no php.net release.
-
3.
The Decade-Old Bug Is Now a One-Week Find
Wiz says php2shell's code sat in PHP for more than a decade and one researcher found it in a week with commercial models; pwn.ai found PwnEscape in three days for under $700 in tokens. CVE disclosures are up roughly 85% while known exploitation rose about 14%. A nine-item checklist for the old internet-facing code nobody owns.
-
4.
OpenSSH sshhole: What to Check on Alpine Linux
OpenSSH 10.6p2 shipped on November 5, 2026, one day after the report, to work around a password bypass on musl Linux with no PAM: an account whose shadow field is exactly * accepts any password of 257 to 1024 bytes. Default Alpine blocks the root case with PermitRootLogin prohibit-password. A five-command check, the fix, and why the maintainers now call long embargoes dangerous.
-
5.
Java4Shell: What to Do Before Your JDK Patch Lands
Java4Shell, CVE-2026-114938, is an out-of-bounds write in HotSpot's x86-64 Base64 intrinsic, scored 8.1 by Oracle, 9.2 by Calif and 9.8 by Red Hat. It needs AVX-512 VBMI and JDK 17.0.6 or later. On November 18, 2026, four of five common image tags still shipped an unfixed JDK. The order: check the CPU, find every JVM, set -XX:-UseBASE64Intrinsics, restart, patch when the vendor ships. -XX:UseAVX=2 stops an ARM64 JVM from starting.
-
6.
Why Auto-Filled Security Questionnaires Kill AI Deals
Auto-filling two hundred questionnaire fields reads as risk, not progress: one clarifying question on a live call collapses an answer nobody on your team wrote. SOC 2 answers maybe sixty of those questions and is silent on the dozen the deal actually hinges on: prompt injection, tenant isolation, what each model provider's terms say about retention and training.
-
7.
The AI Questions on a Security Questionnaire, Answered
The AI section of a security questionnaire in nine themes, from training on customer data to AI incident response. For each: what the reviewer checks, an answer with four parts (status, evidence, owner, and a dated interim step for any gap), and the answer that gets caught when they ask for the document.
-
8.
What Cyber Insurance Underwriters Ask AI Startups
Premiums run roughly $1,500–1,800 a year for $1M in coverage. The application is the exposure: Travelers got its insured's entire policy rescinded after a ransomware claim revealed the MFA attestation wasn't true. ISO issued three generative-AI exclusion endorsements in January 2026. Find that language in your quote.
-
9.
vCISO Math for AI Founders: Why 5 Hours a Month Beats a Full-Time Hire
Full-time CISO costs $200–400K and is over-leveled for pre-Series-A work. The market rate for a fractional vCISO is $2–4K/month. The graduation criteria are explicit.
-
10.
Security Engineer or vCISO? Your First Hire, by Stage
A $220K security engineer dropped into a 12-person startup invents their own job while the enterprise questionnaire in your pipeline sits untouched. Pre-seed it's hygiene, not a role. Seed is 15–25 real hours a month of decisions, so fractional. The first security engineer earns the seat at Series A, when the work flips from decisions to engineering.
-
11.
Most Security Controls Assume a Team You Don't Have
Checklists only grow because nobody who writes them pays the staffing cost. Four controls (prompt-provenance logs, split workforce/customer environments, formal secure-prompting training, a pre-vetted building-block catalog) turn from theater to real at a specific headcount (20 engineers for review discipline, 30 people for training to stop being one senior correcting a junior in a PR comment), not on adoption day. The three-name test for telling a real control from a paper one: the person who runs it, the artifact it produces, what happens when it fails.
-
12.
The MVP Shortcuts That Turn Into Security Incidents
The ship-in-24-hours playbook is right that dark mode, a logo, git branches and new frameworks can wait. Five cuts on the same skip list carry a security job: no account recovery (support-desk takeover), no account deletion (hand-deleted prod rows on the first GDPR Article 17 request), no tests on the payment path, secrets that came with the starter kit, and open admin or debug routes. Each has a cheap minimum fix, and the one-hour pre-launch pass covers all five.
-
13.
Shadow AI at a Startup That Can't Buy a CASB
A CASB is the right tool at 500 employees and the wrong conversation at 12. Three free things instead, inside two weeks: a card-statement-and-SSO-log audit of what's actually connected, a no-blame survey for what isn't, and a one-page data-classification rule. You outgrow it around 25-30 people with no security owner.
-
14.
Model Routers Are Vendors. Review Them Like One.
Anthropic's September 2026 threat report documents three PRC labs harvesting their own users' sessions into Claude (Moonshot's 300,000 requests through 5,380 fraud accounts, Xiaomi's 400,000 through 1,500), with many of those users arriving through the third-party model routers US and European teams use, and the sessions holding pharma capex forecasts and a developer's live Telegram, Feishu and Notion keys. Three things that change on Monday: inventory the endpoints engineers actually hit, no PRC-lab endpoint on customer or company data, and review the router as a vendor.
-
15.
Superintelligence Ban: What It Means for AI Startups
The Ban Artificial Superintelligence Act would pause advanced AI systems trained on 10^25 or more operations until a new Department of AI sets rules, with up to 20 years in prison for individuals. A startup that buys models trains nowhere near that; its exposure is the frontier vendors it depends on. Whether already-deployed models or API access would be paused is still unanswered. A five-item vendor-continuity checklist.
-
16.
The Agent Evidence Pack for a Security Review
The two artifacts an enterprise reviewer actually wants: a tool allow-list carrying what each tool reaches, under which credential, at what risk class, whether it needs confirmation, and who approved it when, plus a tool-call log record with timestamp, actor and actor type, tenant, tool, arguments, outcome, human confirmation, and provider request ID. The trick that keeps both honest is generating the allow-list from the tool registry, because a hand-maintained one describes an agent that no longer exists.
-
17.
Building an LLM Audit Trail That Passes SOC 2
An append-only Ecto schema for LLM calls, and the fields that make a row defensible rather than decorative: the provider request ID is what lets a reviewer check your log against a source you don't control. CC7.2 covers monitoring for anomalies, CC7.3 evaluating those events as incidents, CC6.1/CC6.3 who was authorized to make the call. Retention only binds if your report scopes Confidentiality or Privacy (C1.1 and P4.2), which most Security-only startups never scope, and that is why the prompt table nobody owns is five years old.
-
18.
Claude Code and SOC 2: The Evidence Your Auditor Wants
An eight-row control map for engineers on Claude Code at a SOC 2 Type II company: each row pairs a managed setting with a criterion and the artifact an auditor samples. Managed settings outrank every other source and /status proves it per machine; tool_decision events record whether config, a hook, or a person approved each tool call. Two locks, not one: disableBypassPermissionsMode leaves auto mode on. The gaps are written policy, not software: CC8.1 change control lives in the repo, and file-based settings never reach cloud sessions.
-
19.
How I'd Run Security at an AI-Native Company in 2026
The four-layer stack: prompt injection defense, agent credentials, secrets handling, audit logging. None of these come for free in classic appsec.
-
20.
Prompt Injection Defense for a 5-Person AI Startup
EchoLeak (CVE-2025-32711, CVSS 9.3) exfiltrated Microsoft 365 Copilot data zero-click and walked straight past Microsoft's own cross-prompt injection classifier. The deterministic answer is about a week of work, most of it deletion: every tool scoped to the session's user, outputs validated and not just inputs, a human gate on anything state-changing.
-
21.
Secrets Management for AI Agents on Small Teams
GitGuardian counted 28.6 million secrets exposed in public GitHub commits in 2025, over 24,000 of them in MCP configuration files. The model never needs the key: a broker between the model and every side effect, least privilege scoped to the tool rather than the agent, and one kill switch you've actually tested. Includes what not to buy yet.
-
22.
Stop Sourcing .env: Load It With direnv Instead
Sourcing a cloned repo's .env runs it as shell code: a $(touch pwned) line ran on the spot. Parsing instead of sourcing wasn't enough either. PROMPT_COMMAND runs at every prompt, LD_PRELOAD and GIT_SSH_COMMAND execute too, and zsh's lowercase path walked past the denylist. Four commits in one morning ended with the function deleted for direnv, which loads nothing until you allow it. The caveat: tested on direnv 2.38.2, an edited .env reloads with no new prompt, only .envrc edits are re-approved.
-
23.
Multi-Tenant AI SaaS in Phoenix: Isolation That Holds
Storage scoped correctly and the retrieval read not: one missing where clause, and top-k similarity returns the nearest chunks from every tenant, with a clean log and a passing test suite. Covers shared versus per-tenant provider keys (OpenAI enforces rate limits at the organization and project level, not per raw API key), one scoped queryable with no unscoped path into it, and the CI test that fails the day someone adds one.
-
24.
Detection Got Cheap. Remediation Didn't.
Microsoft's MDASH harness runs 100+ agents and scored 96% on CyberGym at a 50% cost saving over their current offering, and changes nothing about your capacity to fix what it finds. At 12-40 people the constraints are engineer time and the political cost of blocking a release. Write the three-tier rubric before the next scan runs, give the queue one owner, budget fix capacity at 10% of engineering time.
-
25.
What a 96% Security Benchmark Actually Measures
CyberGym is 1,507 historical bugs already patched in the public record, and the agent is handed the pre-patch codebase plus a textual description of the vulnerability: that's reproduction, not discovery. The evidence worth weighing is the 16 new vulnerabilities in the Windows networking and authentication stack and the held-out driver with 21 injected flaws, all identified with zero false positives. Five questions to ask before the vendor call.
-
26.
What 90 Days of a Fractional Security Engagement Actually Looks Like
The week-by-week composite of a typical 90-day fractional CISO engagement. $20–30K in auditor and tooling costs, plus the retainer. What graduation looks like at month four.
-
27.
Migrating 225K Users from AWS Cognito to Auth0 Without Forcing a Single Logout
Production identity migration without forcing 100K users to re-authenticate. Hash incompatibility, lazy migration, MFA token handling: the implementation details vendor docs skip.
-
28.
Surviving Technical Due Diligence as an AI Founder
What the investor's technical reviewer opens first, the AI-specific flags that re-price rounds in 2026, and the gaps you should deliberately leave alone.
-
29.
Your Personal Brand Is an Attack Surface
Every talk, bio, and commit a founder publishes to build a name is also reconnaissance and impersonation fuel. The FBI's 2025 IC3 report logged $3.05 billion in business email compromise losses and complaints mentioning AI carrying more than $893 million; Arup lost about $25 million to a deepfake video call with a fake CFO, while LastPass and WPP stopped the same attack because someone noticed the channel was wrong. The fix is process, not silence: phishing-resistant auth on every profile, one page listing the real accounts, and out-of-band verification for money and access.
Other topic guides
- AI engineering What production AI engineering actually looks like in 2026: the autonomy ladder for agents, the workflow shift, telemetry, and team sizing.
- Engineering leadership Engineering leadership at startup scale: hiring from one engineer to fifteen, rituals that work at small teams, the staff-engineer interview loop.
- Elixir and the BEAM for AI systems Why language choice matters for AI systems: BEAM concurrency for agents, what Go frameworks cost you, and why Elixir is the language AI writes best.
- Omarchy Running Omarchy with coding agents, starting on an M1 MacBook: the install, what breaks, the keyboard, and what a security lead inherits from agents.