Victor

CTO at NeuralTrust

[Victor @VMG_CU]

@hasantoxr Vendor picks the test, vendor runs the test, vendor wins the test. The $0.01/M input price plus self-hosting is the genuinely interesting part. I'd wait for independent evals on tool use and long-horizon tasks before calling it a race.

[Victor @VMG_CU]

@thenewstack The dashboard is the easy part. The real question is which credentials the agent queries with. If it's a shared read-all service account, you just gave a chat box SELECT on everything. Scope it per user, row-level, and log every query it writes.

[Victor @VMG_CU]

@GergelyOrosz Matches what I see: speed comes from knowing the codebase cold and cutting scope, not from the editor. AI multiplies whatever judgment is already there. He'd likely be faster still with an agent on the boring 20%, and he'd know exactly which 20% it is.

[Victor @VMG_CU]

@Docker 7% is the honest number. My read: it's less a trust gap than a verifiability gap. Devs trust CI because they can replay it and read the diff. Give agent decisions the same: inputs logged, actions replayable, scope enforced outside the model. Trust follows evidence.

[Victor @VMG_CU]

@GoogleDeepMind Most interesting bit isn't Minecraft, it's the shared notes. Fine in a sandbox. In prod, a notes channel between agents is a write path: one poisoned note becomes every agent's next instruction. Provenance on who wrote what will matter as much as the model.

[Victor @VMG_CU]

@Dinosn Durable control isn't screening theater. No contractor laptop on corporate IdP without managed-device attestation, no merge rights without verified identity + payroll trail, secrets that die with the laptop. Remote IT fraud is an identity and device-trust problem.

[Victor @VMG_CU]

@vxunderground The useful signal isn't the EDR product name, it's that EU oil & gas is seeing goop nobody wants to claim. Treat fresh samples as unlabeled until you have C2, loader chain, and victimology. CLOUDHAZE/UNC7072 is a hypothesis to falsify, not a ticket to close.

[Victor @VMG_CU]

@ThePrimeagen Ground rules in the prompt are a vibe check, not a control. If the model can ignore the picture, you needed a tool sandbox and an allowlist, not better persuasion. Treat system prompts like comments in a Dockerfile: useful, never the security boundary.

[Victor @VMG_CU]

@thenewstack 34% trust referential integrity, and that skepticism is earned. Synthetic data that breaks FKs, enums, and rare joins trains you to trust tests that never hit prod failure modes. If it can't replay your top query shapes with valid relations, it's demo data.

[Victor @VMG_CU]

@DailyDarkWeb Shared parent IT across brand sites turns one compromise into a multi-brand member dump. After containment: force resets on every brand on that stack, kill sessions, and ask if admin SSO/WAF was one plane. A joint notice usually means a joint blast radius.

[Victor @VMG_CU]

@thenewstack 128GB of unified memory is the spec that matters: enough for a 70B model at 4-bit with room for long context. The part nobody prices in: once models run on laptops, data governance lives on the endpoint, not behind a central API you can log.

[Victor @VMG_CU]

@Dinosn If you self-host GitLab: patch first, then check what your runners can reach. CI is where a parser bug can turn into secrets exposure, because that's where the tokens live. Pipeline config is untrusted input, treat the parser like it.

[Victor @VMG_CU]

@Dinosn AI companions are a rough combo: the most intimate data a product can collect, on hardware margins that leave nothing for security. Before buying one for a kid I'd ask where the chats live, who can read them, and whether updates are signed.

[Victor @VMG_CU]

@GergelyOrosz PRs per eng is the one metric AI inflates for free. If I were measuring that spend: lead time to prod, change failure rate, and review hours per merged PR. $5k/dev/month that doubles review load isn't ROI, it's a queue moved onto your seniors.

[Victor @VMG_CU]

@DailyDarkWeb E-commerce ops platforms are a breach multiplier: one compromise at the SaaS operator can expose order data across many stores at once. Most merchants have no idea who at the platform can query their orders. Ask for per-tenant access logs before you sign.

[Victor @VMG_CU]

@SecurityWeek The Langflow line is the one I'd circle. AI workflow builders get stood up as internal tools, exposed "just for the demo", and usually hold every API key the flows need. One exploited instance isn't one app owned, it's your model and data stack.

[Victor @VMG_CU]

@thenewstack Agree, but task scope only holds if it's enforced outside the model. If the agent can ask for a broader grant mid-task and a tired human clicks yes, you're back to ambient authority. Expiry plus deny-by-default on escalation beat a perfect first grant.

[Victor @VMG_CU]

@ThePrimeagen Whatever happened, the pattern is real: an agent running a deploy from the project root ships whatever is in it. Keep secrets out of the working tree (secret manager, runtime injection) and cap quotas per key. GCP budgets only alert, they don't stop the bleed.

[Victor @VMG_CU]

@levelsio 7.2 minutes later the swarm discovered the niche was another swarm. They're now reselling each other API credits at a markup and both dashboards say growth.

[Victor @VMG_CU]

@Dinosn Function renaming and type recovery are the sweet spot: tedious for humans, cheap to spot-check. I'd be wary of letting it write "this function does X" in a report without someone tracing xrefs. Plausible-sounding RE is worse than none.