Jascha

Technologist, #infosec, #FOSS, #Linux, #OSINT #AISec #CyberSec
Founder of and ≠ Endorsement

[Jascha @jascha]

@VolvoCarUSA Can never get an email followup from your support. Any other avenues than that?

[Jascha @jascha]

The eBook: https://t.co/6drPwjZHWx

[Jascha @jascha]

Anthropic has validated what we've been working on for 2 years with their Zero Trust eBook for AI Agents. @ThirdKeyAI and the https://t.co/WIZ8u0he7x even fill in the gaps.

[Jascha @jascha]

@claudeai All I get for $200 month now...

API Error: 400 messages.27.content.11: `thinking` or `redacted_thinking` blocks in the latest assistant message cannot be modified.
These blocks must remain as they were in the original response.

[Jascha @jascha]

Google announced that all Android app developers must register centrally, pay a fee, and submit government ID, or their apps will be blocked on every device. over 67 organizations oppose this. https://t.co/cJlLZL7ThS @AlteredDeal #KeepAndroidOpen

[Jascha @jascha]

VectorSmuggle is on HNS. Steganographic exfiltration via vector embeddings, plus a cryptographic defense (VectorPin).
https://t.co/NHB4FSUUpm

[Jascha @jascha]

Small-angle orthogonal rotation hides arbitrary payload in embeddings, undetected by off-the-shelf statistical defenses across every model-corpus pair we tested.
arXiv: https://t.co/A1NvFg44LX

[Jascha @jascha]

@TheWhizzAI https://t.co/GilVXV5L2g and https://t.co/pG2i0jZRkY address these issues.

[Jascha @jascha]

RAG security today is roughly where cloud storage was before people realized S3 buckets were public.

[Jascha @jascha]

Added a visual explainer for VectorSmuggle: Steganographic exfiltration in embedding stores

https://t.co/P9sltNol94

[Jascha @jascha]

Codebase with Jupyter demos:
https://t.co/KjuTCwY2we

[Jascha @jascha]

I published a new security preprint:

VectorSmuggle: Steganographic Exfiltration in Embedding Stores and a Cryptographic Provenance Defense

The paper looks at an integrity gap in RAG infrastructure: embeddings are usually treated as opaque numerical artifacts, but vector stores generally lack native/default controls for embedding integrity or provenance.

The finding is nuanced: simple anomaly detectors catch many crude perturbations, but small-angle orthogonal rotation exposes a harder class of embedding tampering. Statistical detection helps, but it should not be the load-bearing control.

I propose VectorPin: a minimal cryptographic provenance protocol that signs a canonical commitment over the source, model, and vector bytes. If the vector is modified after pinning, verification fails.

Paper:
https://t.co/tPVU1is3zT

https://t.co/liGCeeb36t

[Jascha @jascha]

@BBGreatMoments That was not "safe"

[Jascha @jascha]

@akaclandestine This my project which can be found here:
https://t.co/KjuTCwY2we

[Jascha @jascha]

New paper on how Symbiont uses the ORGA loop to secure agents.

"Typestate-Enforced Agent Loops: Making Policy Gates Unskippable at Compile Time"

https://t.co/VcSF8prJy7

[Jascha @jascha]

@lifeof_jer Issues like this is exactly why I built Symbiont.

[Jascha @jascha]

Symbiont is a full implementation of OATS under Apache 2.0 license:
https://t.co/NZIDiYaEt6

[Jascha @jascha]

AI agents are getting more powerful.

The trust layer around them is not.

Today, too much agent safety still depends on prompts, wrappers, and best-effort guardrails. That is not enough for systems that can actually take action.

Introducing OATS: the Open Agent Trust Stack.

OATS is an open specification for zero-trust AI agent execution built around tool contracts, identity, policy, and auditability.

It is also grounded in real implementation work. Symbiont has been applying these ideas in practice over the past year.

The goal: make safe behavior enforceable by design, not optional at runtime.

https://t.co/RBYFYOZdGJ

#AI #AISecurity #AgenticAI #OpenSource

[Jascha @jascha]

@icanvardar Like that guy from 50 first dates.

[Jascha @jascha]

@infosec_fox Inside a cat poop inside the cat box.