SShortSingh.
TopTechnologyAI & MLStartupsProgrammingScienceBusinessFinanceCrypto & Web3World
Back to feed
Programming

Study finds NVIDIA OpenShell's audit export empty; developer builds independent evidence ledger

0
·1 views

An independent evaluation of NVIDIA's OpenShell, a sandbox runtime for AI agents, found that its blocking worked but its evidence trail did not. A blocked DNS query left no record, and the OCSF audit export produced zero records in test runs. The author responded by building Agent Seatbelt, a vendor-neutral policy and evidence layer that sits above existing agent runtimes. It uses externally run policy decisions and an append-only, hash-chained ledger that can be verified offline. In testing on a real runner, the layer reconciled 13 of 13 actions against independent ground truth.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingDEV Community ·

DREX: Brazil's central bank digital currency faces privacy questions

The Central Bank of Brazil is developing DREX, a digital currency built on a permissioned Ethereum-based network called Hyperledger Besu. Unlike decentralized cryptocurrencies, it is run by authorized banks and institutions, with the central bank retaining oversight of the system. Pilot tests revealed that, without extra safeguards, transactions could be visible to network participants, and researchers reportedly exposed balances and transfers during one trial. The bank is evaluating privacy technologies such as zero-knowledge proofs and anonymous transfer protocols, though full privacy is seen as incompatible with anti-money-laundering rules. Implementation is underway, with expansion expected through 2025 and 2026.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

OpenClaw's heartbeat feature: cheap local checks first, LLMs only when needed

OpenClaw, formerly known as Moltbot and Clawdbot, has gained traction for its local automation that monitors a machine, spots drift, and acts early. Its heartbeat feature runs periodic ticks through a layered pipeline, starting with inexpensive deterministic checks such as process status, file state, queue depth, and API health. Model escalation happens only when those checks leave real ambiguity, which reduces cost and unpredictability. The article presents the heartbeat as a control-plane primitive for agent infrastructure, organized as a five-stage loop with strict probe schemas, TTL-based caching, and a health state machine using hysteresis. The guidance is aimed at developers building their own agent systems who want predictable behavior and fewer unnecessary model calls.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Origin Data Lab Examines Stereo Camera Calibration Limits for Physical AI Data

Origin Data Lab has been evaluating the geometric consistency and calibration limitations of its stereo capture workflow, which is used for Physical AI and robotics data. The company's Technical Field Note #04 argues that a single average error value cannot fully describe calibration quality, and that the distribution of epipolar errors can reveal residual misalignment. It also highlights that calibration parameters must match the actual capture setup, since changes in resolution, cropping, or camera mode can affect their validity. Further points include evaluating errors across field angles and using appropriate projection models for wide-angle and fisheye lenses. The note concludes that reliable real-world datasets must be measured, verified, and documented rather than simply collected.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Armory Castle exploits BeyondTrust Remote Support flaw CVE-2026-102764 in appliances

Google's threat intelligence group and Mandiant report that a suspected China-nexus actor tracked as Armory Castle has exploited CVE-2026-102764 in BeyondTrust Remote Support appliances since late August 2026. BeyondTrust rates the authentication bypass 9.3 and says it affects Remote Support 26.2.3 and earlier versions. The vendor patched SaaS customers on 8 October 2026 and disclosed the flaw on 15 October, while self-hosted appliances require manual patching unless automatic updates are enabled. According to Mandiant, the attacker used the flaw for initial access, escalated to root, deployed a webshell, and extracted stored credentials to reach downstream systems. Researchers also note that search-engine fingerprint coverage for the product is incomplete, so exposure counts can be misleading in either direction.

0 comments Read more at DEV Community