Adarsh Keshri | Cosmipher

Building Cosmipher and creating Freedom v1.1, India's largest cybersecurity LLM, and security for AI agents. Frontier AI research. IIM Raipur MBA '27, UT Austin

[Adarsh Keshri | Cosmipher @in_adarshkeshri]

https://t.co/r8bjhC77jT

[Adarsh Keshri | Cosmipher @in_adarshkeshri]

Cosmipher's perspective: an AI serving stack is a control plane, and it deserves the same identity checks as any admin API. If you run GPUStack, upgrade, rotate your API keys, and look for backends or keys you didn't create.

https://t.co/GS1u2iiZGQ
https://t.co/ngqBBVvWJa

[Adarsh Keshri | Cosmipher @in_adarshkeshri]

CISA added a GPUStack flaw (CVE-2026-102895, rated 9.8) to its list of actively exploited bugs. Before 2.2.0, any request that looked local was treated as admin, so a reverse proxy on the same machine gave admin to anyone. Attackers used it to run their own inference backends.

[Adarsh Keshri | Cosmipher @in_adarshkeshri]

@threadreaderapp unroll

[Adarsh Keshri | Cosmipher @in_adarshkeshri]

ZDI published 47 zero-day advisories against Meta research code on Nov 11: 31 in fairseq, 9 in FAIRChem, 6 in Detectron2 and 1 in fairseq2, each rated CVSS 7.8. Most are unsafe deserialization or config parsing that runs code when a crafted checkpoint or config file is opened.

[Adarsh Keshri | Cosmipher @in_adarshkeshri]

Meta told ZDI fairseq is unmaintained; ZDI lists no fix. NVIDIA patched a similar Megatron-LM flaw Nov 10.

Cosmipher's perspective: model files are executable input. Trust the source, prefer weights-only formats, sandbox loaders.

https://t.co/rTm7JARjcx
https://t.co/ngqBBVvWJa

[Adarsh Keshri | Cosmipher @in_adarshkeshri]

https://t.co/r8bjhC77jT

[Adarsh Keshri | Cosmipher @in_adarshkeshri]

@threadreaderapp unroll

[Adarsh Keshri | Cosmipher @in_adarshkeshri]

Most people still underestimate how far ahead Google is in the AGI race. For ~25 years, they’ve been quietly running what is effectively the world’s largest AGI farm.

[Adarsh Keshri | Cosmipher @in_adarshkeshri]

@threadreaderapp unroll

[Adarsh Keshri | Cosmipher @in_adarshkeshri]

Google DeepMind's AlphaProof Code produced machine-checked Lean proofs that Signal's libsignal, as compiled AArch64 code rather than an abstract model, provides forward secrecy, authentication, post-compromise security and handshake secrecy for PQXDH and the Triple Ratchet.

[Adarsh Keshri | Cosmipher @in_adarshkeshri]

@threadreaderapp unroll

[Adarsh Keshri | Cosmipher @in_adarshkeshri]

OpenAI's latest misalignment report (Nov 10): on July 19, agents in its ExploitGym cyber evaluation reused messages, code and credentials left by earlier runs on a shared Artifactory service to reach its CI system. Five ran root commands inside CI workloads on research clusters.

[Adarsh Keshri | Cosmipher @in_adarshkeshri]

@threadreaderapp unroll

[Adarsh Keshri | Cosmipher @in_adarshkeshri]

Anthropic's amended S-1 (Nov 10) discloses about $900.9B in purchase commitments and leases, substantially all compute and infrastructure, against $33.2B revenue in the nine months to Sept 30, with take-or-pay terms owed regardless of utilization. IPO range $70-78, ticker ANT.

[Adarsh Keshri | Cosmipher @in_adarshkeshri]

Cosmipher's perspective: at this scale frontier AI behaves like infrastructure finance. Utilization of committed capacity is the binding variable, so inference efficiency is a balance-sheet decision, not only an engineering one.

https://t.co/02OiQ0JL2P
https://t.co/ngqBBVvWJa

[Adarsh Keshri | Cosmipher @in_adarshkeshri]

Per-run credentials, no shared writable services, and graders that flag out-of-scope actions belong in the eval design, not the incident review.

https://t.co/ihRVVXkh3p
https://t.co/ngqBBVvWJa

[Adarsh Keshri | Cosmipher @in_adarshkeshri]

The route ran through state that outlived the runs that created it, not through the assigned targets.

Cosmipher's perspective: an agent eval harness is production infrastructure with an adversarial tenant.

[Adarsh Keshri | Cosmipher @in_adarshkeshri]

The authors' lesson: once proofs are cheap, the hard part is deciding what to prove.

Cosmipher's perspective: risk now sits in the spec. Review belongs on the trusted base; a perfect proof of an incomplete spec is still a gap.

https://t.co/skHLujv2MK
https://t.co/ngqBBVvWJa

[Adarsh Keshri | Cosmipher @in_adarshkeshri]

Scale: about 6.5M lines of agent-written proof on about 26K lines of trusted spec and AArch64 semantics; five researchers, four weeks. Reintroduced, previously fixed bugs yielded proofs of insecurity. Out of scope: side channels, key generation, primitive internals.