SShortSingh.
TopTechnologyAI & MLStartupsProgrammingScienceBusinessFinanceCrypto & Web3World
Back to feed
Programming

vLLM Nemotron processor bug disabled Pillow's decompression-bomb guard, enabling memory exhaustion

0
·2 views

vLLM has patched CVE-2026-116177, a high-severity flaw affecting versions 0.10.2 through 0.24.x that can be exploited without authentication. The vulnerability stems from Nemotron-VL multimodal processors globally setting Pillow's MAX_IMAGE_PIXELS to None, which removes the library's safeguard against decompression-bomb images. A maliciously crafted image sent to the chat completions endpoint can therefore exhaust memory and crash the worker process. A researcher testing Pillow 10.2.0 showed that tiny header-only PNG files were refused under default limits but accepted without warning once the limit was disabled. The analysis also found that small compressed PNGs could expand into hundreds of megabytes of memory when decoded.

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 ·

Single Large AI Model or Several Specialized Ones? A Cost, Latency, and Accuracy Tradeoff

A developer weighing how to architect an AI system faces a choice between sending all requests to one frontier model or routing tasks across several models. The article argues the single-model approach offers simplicity, broad capability, and low upfront engineering cost, which suits early-stage products with uncertain traffic. As volume grows, it contends that routing routine work to smaller models can sharply reduce costs and latency, since frontier models are considerably more expensive per token and slower. It also notes that fine-tuned small models can sometimes outperform large general models on narrow tasks. The author concludes the right choice depends on the workload and on weighing cost, latency, accuracy, and maintenance.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Python Parity Validator Gave False PASS on Nested JSON Type Mismatch

An audit uncovered a flaw in a replay-versus-live trace comparator built to verify decision pipelines, where nested JSON lists and dictionaries were compared with ordinary Python equality. As a result, a value of 1 in a list was treated as equal to True, causing a strict PASS for a recorded type difference. The bug was reproduced from a preserved source snapshot, and a public synthetic fixture now correctly reports a FAIL with the exact divergent leaf identified. The defect was not caught by the 216 existing tests, which covered scalar numeric semantics but not the same behavior at every JSON leaf. The project's author describes the fix as recursive and type-aware, and notes that CSV equality remains deliberately unverified without provenance.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

AI coding agents speed up code production, but integration and review now slow delivery

Coding agents can produce corrections, tests and API changes quickly, yet someone still has to verify that these changes work together and can be shipped. GitHub's Octoverse 2025 report cites an average of 43.2 million merged pull requests per month, up 23% year over year, though it does not isolate the share driven by AI. Anthropic reportedly grew from about 2,500 to over 30,000 integrated changes per day between May and September 2026, and its October 2026 account of its Scion project describes reworking version management to handle that volume. The article argues that review capacity and continuous integration pipelines, rather than code generation, now set the pace of delivery. It also warns that independently passing changes can conflict once combined, so merging without textual conflicts still requires testing the combined result.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Mac dictation apps may send transcripts to the cloud even when speech is processed locally

A Mac dictation tool can transcribe speech on the device and then forward the resulting text to a cloud model for cleanup, so the workflow is not fully local. Speech recognition turns microphone audio into text, while cleanup edits that text by removing fillers and applying spoken corrections. Each stage may run in a different place, and the cleanup service receives only text, though that text can still contain sensitive details from the recording. Users should count every recipient, including relay servers and other machines on their network, and check whether a fallback to cloud processing could occur silently. For sensitive work, the article recommends choosing a setup that clearly stops when local processing is unavailable.

0 comments Read more at DEV Community