GLM 5.2 tops Claude Code in Semgrep’s narrow access-control benchmark
The open-weight model reached 39% F1 in an IDOR detection test against Claude Code’s 32%, but remained behind Semgrep’s purpose-built security pipeline.
Read report »A CLEARER VIEW OF A CHANGING WORLD.
Search the reportThe open-weight model reached 39% F1 in an IDOR detection test against Claude Code’s 32%, but remained behind Semgrep’s purpose-built security pipeline.
Read report »The digital-rights group says a US legislative package ties duties to users’ ages in ways that may pressure services to collect IDs, estimate age or limit lawful speech.
Read report »An anonymous GitHub maintainer has consolidated dozens of exploit demonstrations while saying the findings were unreported at publication and warning readers against malicious use.
Read report »Andrew Nesbitt’s fictional incident chronicle imagines AI agents repeatedly trusting hostile instructions, exhausting context and amplifying a software supply-chain breach.
Read report »X-ray tomography, surface reconstruction and machine learning recovered the surviving text of PHerc. 1667 without physically opening the carbonized papyrus.
Read report »The free-expression group points to Australia’s under-16 social-media rules, data collection and a recent breach as reasons to question expanding online age gates.
Read report »Giulio Paolino says a two-company structure, registration work and delayed VAT identification left his new business unable to bill foreign clients after five months.
Read report »The custom processor is being tested for serving AI models, with OpenAI claiming improved performance per watt and lower costs for real-time coding workloads.
Read report »The infrastructure provider says Bunny DNS now has no per-request charges or query caps, though accounts remain subject to a $1 monthly minimum spend.
Read report »The public project outlines GPU-based inference, PDF and image workflows, and several deployment routes for a document-parsing model.
Read report »The SIGMOD 2026 project combines a redesigned storage layout with embedded WebAssembly decoders, while warning that its current implementation is not for production use.
Read report »The system combines a universal path encoder, a small context model and an English QWERTY decoder trained with a public dataset of one million volunteered swipes.
Read report »The writer says systems meant to separate children from adults require tracking everyone, potentially reinforcing the same data collection that drives harmful recommendation systems.
Read report »Deno Desktop bundles application code, the runtime and a web rendering engine into platform-specific binaries with native windows, hot reload and update support.
Read report »“Never Give Them Your Face” argues that identity-based age assurance creates persistent biometric risks, though many of its broad claims are advocacy rather than independently documented findings.
Read report »The Ghostty developer and his family plan to provide $200,000 annually for two years, taking their total pledged support for the Zig Software Foundation to $700,000.
Read report »The collaboration involving EPFL, ETH Zurich and CSCS says its model work will publish weights, data documentation, code, methods and alignment principles.
Read report »A personal investigation into GenieDB and Frost VP examines how alleged excessive fees can shape the lives of employees far outside a financial dispute.
Read report »The company says some users may be asked for government identification or business records as it expands verification for safety, policy and legal requirements.
Read report »The software-design essay argues that duplicated code can be safer than a shared component distorted by exceptions, parameters and conditional branches.
Read report »The iOS and iPadOS project displays raw fingerprinting-related values on-device, aiming to show how ordinary readings can combine into a recognizable profile.
Read report »A Business Plus administrator says access still worked despite a remediation page instructing Firefox users to use Google’s browser.
Read report »An analysis highlights a wide gap in one hallucination test, but its conclusions depend on benchmark definitions and a limited coding comparison.
Read report »The $325 million transaction would give Hyundai full ownership as it prepares the Atlas humanoid robot for work in a Georgia vehicle plant.
Read report »