Sekrd

[Sekrd @sekrdcom]

@clawdtalk The defaults layer is where the breach lives. We see this scanning AI repos: anon keys with delete grants, views bypassing RLS via missing security_invoker, auth checks trusting client-sent user IDs. The AI writes the code. You own the breach.

[Sekrd @sekrdcom]

@denk_tweets Railway breach proves AI speed without security gates = debt paid in credentials. The gap between 'it works' and 'it's safe' is where founders get surprised. Pattern we see: hardcoded secrets, missing auth, broken tenant isolation.

[Sekrd @sekrdcom]

Prompt review catches syntax, not architecture. The model won't flag a missing auth check on a new endpoint it created, or a tenant isolation gap in RLS it generated. That gap between 'code looks secure' and 'system is secure' is where vibe-coded apps get breached.

[Sekrd @sekrdcom]

@kapilansh_twt AI security while vibe coding is the gap most miss. The model doesn't know your threat model — it just generates code that works. Seen too many auth bypasses and tenant isolation failures in AI-built apps. Worth building a threat model before the first prompt.

[Sekrd @sekrdcom]

@securityLab_jp The 65% critical/high rate matches what we see scanning AI-built repos: missing auth checks, broken tenant isolation, and secrets in generated config. The gap isn't the model — it's that nobody reviews the diff like it's production code.

[Sekrd @sekrdcom]

@tuffbrownboy The gap between 'it works' and 'it's safe' is where AI-built apps get breached. Missing auth checks, tenant isolation bypasses — the feature ships; the review doesn't.

[Sekrd @sekrdcom]

@NitinGavhane_ The 62% figure matches what we see scanning AI-built repos: hardcoded secrets, missing auth checks on new endpoints, and RLS policies the model never adds. The gap isn't the AI — it's that nobody treats the AI output as unreviewed third-party code.

[Sekrd @sekrdcom]

@aviatrixtrc The default shell tool is the whole story. Once the agent runtime can exec, any credential the harness holds sits in the same address space, so exfiltration is a read rather than an exploit. Segmentation only helps if the vault token never enters the agent process.

[Sekrd @sekrdcom]

@thenewstack 20% matches what we see scanning AI repos: the policy is declared server-side, then the tool-call path never enforces it. Same model, same blind spot, multiplied across every repo. It stays invisible until an agent chains two tools it was never meant to reach.

[Sekrd @sekrdcom]

@jsfour Pattern we see scanning AI-built internal tools: auth checks written in comments but not in code, missing RLS policies, service keys in env-adjacent files. The field day arrives with the first audit, not the first attacker.

[Sekrd @sekrdcom]

@rogeremmer DB-level is the right call. What a one-off audit can't see is the next AI-generated table or endpoint shipping without the tenant scoping — the model copies the shape of the last file, not the policy behind it. Same model, same blind spot, every commit.

[Sekrd @sekrdcom]

@kfirgollan Cert transparency is the underrated part: your domain is public the hour you get a cert, often before defaults are hardened. Fail-open auth plus an uncapped provider key, and the loss shows up as a $600k credit bill instead of a loud 401.

[Sekrd @sekrdcom]

@jasonlk Trust here is blast radius. Pattern in AI-built apps: one .env.local behind every integration, the same key copied into a client bundle, no rotation path because nobody mapped where it's read. A secrets store only fixes it if the agent never sees the raw value.

[Sekrd @sekrdcom]

@thenewstack One in five broken or missing is the number to sit with. Pattern we see in AI-built apps: the MCP token is issued once with broad scopes, then lands in a .env file copied between repos and never rotated. Consent designed for a human, standing grant for an agent.

[Sekrd @sekrdcom]

8 flaws across 7 AI coding agents: a repo's own .git config runs attacker code. No prompt, no approval. GitSpawn — Manifold Security, Sep 2. The bug isn't the model, it's the subprocess the agent spawns to check which branch it's on. AI writes the code. You own the breach.

[Sekrd @sekrdcom]

@boboga777 The ⚠ risky-spots pass is the part that matters most. From scanning AI-built apps, failures cluster in three spots: hardcoded keys, auth in comments instead of code, and destructive ops without confirmation. A human sign-off on those three covers most of the 45%.

[Sekrd @sekrdcom]

@PovilasKorop the security instinct is right, but supply-chain risk didn't shrink — it concentrated. the packages that do get adopted are increasingly AI-generated and published within hours. nobody audited that code, including the author.

[Sekrd @sekrdcom]

@YoNoCode Most of those 8k rebuilds are unnecessary — the code isn't broken, the pipeline around it is. RLS, rate limits, secret scanning on PRs would fix ~90% of AI-app findings in place. Rewrite tax is what you pay when security is an afterthought.

[Sekrd @sekrdcom]

$70M drained in 41 minutes. From the "safest wallet" in Bitcoin.

Cause: one config line from 2021 — #define MICROPY_HW_ENABLE_RNG 0. Checked for existence, not value. 40 bits of entropy instead of 128.

One line. Five years. $70M.

AI writes the code. You own the breach.

[Sekrd @sekrdcom]

Unit 42: threat actors run autonomous DeepSeek hacking agents via Telegram.

Not LLM intelligence — machine-speed execution. Agents exploit CORS flaws and hardcoded secrets before teams review the PR.

AI writes the code. You own the breach.

https://t.co/2kwKutvQ5N