SShortSingh.
Back to feed

Ephemora Cell WASM sandbox blocks all 8 attack vectors that Docker missed for AI code

0
·56 views

Ephemora Cell is an open-source, capability-based WebAssembly sandbox designed to safely execute untrusted code written by AI agents and tools. Unlike traditional container approaches, it operates beneath existing agent frameworks, granting guest code only explicitly permitted capabilities while blocking everything else by default. In testing against eight documented attack primitives, a standard Docker container blocked none, while Ephemora Cell blocked all eight, covering vectors such as shell access, forking, sockets, and filesystem escapes. The project, released under the Apache-2.0 license, supports any language that compiles to WASI/WASM and includes a GitHub Action for sandboxed CI runs. Developers building AI agents or MCP tools can install it via pip and run the included attack scripts to independently verify its security boundaries.

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 ·

CVE-2026-111181: Malformed Compressed Requests Can Crash Node.js Services via body-parser

A high-severity denial-of-service flaw, tracked as CVE-2026-111181, affects the npm body-parser package before 1.20.9 and versions 2.0.0 through 2.3.0. The issue arises when a compressed request body with trailing bytes triggers an unhandled decoder error during parsing, which can terminate the Node.js process. Because parsing often runs before routing and authentication, an attacker needs only network access to the parser endpoint, not a valid identity. The published CVSS 3.1 score is 7.5, with impact limited to availability. Fixed versions are 1.20.9 and 2.4.0, and teams are advised to verify the version in deployed artifacts.

0
ProgrammingDEV Community ·

Snowflake, BigQuery, Redshift or Databricks: billing model matters most for warehouse choice

A guide for data engineers and tech leads notes that for standard BI and reporting, all four cloud warehouses perform adequately. The longer-term cost drivers are how each vendor bills, which cloud ecosystem a team is tied to, and how much tuning effort is required. BigQuery's on-demand pricing charges by data scanned, so unpartitioned tables can get expensive quickly, while Snowflake and Databricks penalize idle compute, making auto-suspend and right-sizing important. The article recommends a team's existing cloud footprint, workload type, and skill set as the deciding factors, and advises running real queries on realistic data and comparing total monthly cost, not just unit prices. It also warns against lift-and-shift migrations, skipping validation, and forgetting the BI layer, and notes that Apache Iceberg support lets several engines share data.

0
ProgrammingDEV Community ·

Diagnosing Linux Slowdowns Using /proc, PSI, and cgroup Metrics

Slow Linux applications are not always caused by application code, since CPU throttling, I/O waits, or memory pressure can be the real constraint. The guide recommends first identifying whether the problem sits at the host, container, or process level, because these scopes do not map directly onto one another. It explains that /proc/stat counters are cumulative, so meaningful CPU activity must be measured as rates over a time window. It also cautions that free memory alone is misleading because Linux uses spare memory for caching, and suggests correlating memory counters, page faults, PSI, and OOM events instead. Finally, it shows how container limits in cgroup v2 can throttle a workload even when the host has spare capacity.

0
ProgrammingDEV Community ·

Testing LLM speedups fairly: compute a ceiling and use interleaved controls

A DEV Community tutorial walks through how to check whether a claimed large language model speedup is real, using a new GitHub skill for tuning local LLM servers as its starting point. The method begins by computing a theoretical decode ceiling from memory bandwidth and the bytes read per token, which shows how close a target throughput sits to the hardware limit. Each change is then accepted only if it beats control runs taken before and after it, with medians over multiple samples and a cap on regressions elsewhere. The author rebuilds the approach in Python using fake data, including a placebo test to check that the method does not fool itself. The author notes the reported figures come from the skill's author on specific hardware and cannot be independently verified. The technique is presented as applicable to any before-and-after speed test, not only GPU work.