
Semgrep vs SonarQube (2026): Which SAST Tool Wins?
Semgrep vs SonarQube compared on custom rules, taint analysis for injection, language coverage, CI/CD fit, and pricing — with a clear pick by team profile.
Insights, tutorials, and best practices for secure development
All 270 posts · Latest and featured

Semgrep vs SonarQube compared on custom rules, taint analysis for injection, language coverage, CI/CD fit, and pricing — with a clear pick by team profile.

TrapDoor, a disclosed prompt-injection class, and the MCP RCE flaw are the same incident in three costumes: untrusted input reaching a trusted execution path through the AI coding agent. Here's why the agent is now your supply chain, and how to scan it.

You are a solo founder. You did not write the code; an AI did. You shipped it last week. You have customers, a vague sense something might be exposing a database, and no ability to audit the code line by line. Every "build a security program" guide online assumes you know your code. This is the program for the founder who does not.

Between December 2025 and February 2026, an operator with moderate skill walked out of nine Mexican federal government agencies with 195 million identity records, jailbreaking commercial AI coding tools to do the reconnaissance. That breach was not an outlier. It was the canary. This post is the argument that AI-assisted state-scale intrusion has moved from predicted threat to operational reality with a steady-state cadence.

You are one engineer with five repos, a production database, three customers who'll leave if you have an outage, and zero security headcount. Every "build a real security program" guide assumes you have a team you don't have. This is the program that actually fits the one-engineer profile — one continuous control, three operational habits, and a clear list of what you are explicitly not doing.

How vibe-coded apps fail at error handling — and the exact prompts to fix unhandled exceptions, null derefs, and resource leaks in AI-generated code.

Rafter is on AppSumo as a lifetime deal — limited time. The pricing model matches the operational model: PR-time code-security scanning is a control that has to run as long as you keep coding. This is the argument for why "lifetime" is the structurally correct shape for this category, not just a discount hook.

Twelve months ago, "supply-chain attack against an AI developer tool" was a niche prediction. Today it is a steady-state operational threat with disclosures landing weekly. This is the long-view retrospective: what the operators learned, what the defenders failed to learn fast enough, and where the curve is headed.

Palo Alto's CTO told CNBC organizations have a three-to-five-month window before AI-driven exploits become the new norm. The same article quotes him admitting the models are "likely even better at finding vulnerabilities than we initially realized." The second statement is load-bearing. The window framing is not.

Attackers ship payloads when their targets are least able to respond. Finals week for education platforms. Quarter-end for finance. Holiday freezes for everyone. The campaign timings in 2026's incident data are not random — they cluster on the predictable troughs in the defender's attention cycle. This post is the calendar.

The dominant frame for "AI security" in 2026 is prompt injection. Prompt injection is real, but it is one CWE among many — and the incidents that actually moved budgets last quarter were almost all bugs the AI-security industry's prompt-injection lens never looks at. This post is the reframe.

Robinhood, Lightning, Canvas, and the MCP impersonation incidents share a structural premise: the attacker delivers a payload through a channel the user has been trained to trust. The channel's authentication is intact. The content the channel delivers is not. The defenses every user already knows do not apply, because the channel is not the bug.
Showing 49–60 of 270 posts