π Fine-tune your own frontier model! π₯π₯
Here's a skill repo to fine-tune Inkling 2 Small. Import it to your coding agent and you're good to go: https://t.co/LSCHr5bSiF
It contains everything your agent needs: a SKILL.md file, Python scripts and a requirements.txt file to get your Python environment set up.
This is a proven recipe, used to fine-tune the Tess-4.5-594B-A17B model. Enjoy! π
Most web hacks start with one value in a URL. If a user can change it, an attacker can too.
Attacker-controlled input, and there's 6 ways it goes wrong:
- IDOR
- Open redirect
- XSS
- SQLi
- Path traversal, and
- SSRF
We're releasing the most capable version of Drost that we can safely put directly in your hands.
Drost Edge runs autonomous security engagements from your own computer.
Give it a web application youβre authorized to test. It maps the surface, forms attack hypotheses, exploits what it finds, and keeps going until it can prove how far an attacker could actually get.
Sign up today, and get $100 in promotional credits: https://t.co/ITR3wJS9jm
Iβm releasing Drost Edge today.
Drost Edge runs autonomous security engagements from your own computer. You can watch every decision, command and attack path as it happens, and your report is ready immediately after the engagement.
A year of R&D, and distilled down to less than 10MB binary. This is also the most capable version of Drost I can put directly in usersβ hands, considering safety.
The public beta is now open for macOS and Linux. Every new signup gets $100 in promotional credits.
Sign up here: https://t.co/qyIChRqW0y
π
@BenjDicken@mbjschirp For the PII point, encryption at rest and query permissions solve different problems. If the agent can SELECT plaintext, disk encryption won't limit what it sees. I'd check its actual database role's row/column access before adding Slack or customer data.
@NathanMcNulty One scope caveat is: GSA's compliant-network check rejects replayed access tokens in near real time for integrated CAE-capable services. Existing app sessions need separate review. Here's from Microsoft's docs: https://t.co/b0gciKat6M
@merill Could the team reach trusted backups, recovery credentials and the runbook with AD unavailable though? I'd put that at the start of the exercise? A timed restore in an isolated lab would make the rebuild vs recover more concrete I'd say..
@mhmazur Can the coding agent edit the deploy gate or write the verdict file? Before widening scope, I'd put those under a separate identity. A second model can review the change.. But it shouldn't share the writer's authority to approve it.
@AikidoSecurity Update for anyone following this: Adobe's Sep 7 hotfix is out. Its recovery guidance also says to rotate exposed integration/payment credentials at their source; changing Commerce's encryption key alone won't invalidate them. Adobe guidance: https://t.co/D8U06KrGQ9
Two days ago, an attacker stole nearly 4,000 BTC ($320M USD) from Liquid Network, a Bitcoin sidechain.
Yesterday, they returned 3,400 BTC, keeping 600 ($47M USD) as their "bug bounty".
We turned this partial refund into a heist musical: βKeep the Change.β
Enjoy!
Weβre Drost. We use AI to test how far an attacker could get in your systems, with people reviewing the work.
The point is to give your team evidence it can use to decide what to fix.
We also make the occasional heist musical. Youβll see what we mean.