HackerNews Digest

The U.S. government issued an export‑control directive requiring Anthropic to suspend all access to its Fable 5 and Mythos 5 models for any foreign national, including Anthropic employees abroad, effective immediately; other Anthropic models remain available. The order, received at 5:21 p.m. ET, cites a national‑security concern about a “jailbreak” method for Fable 5 but provides no specifics. Anthropic reviewed the reported technique, finding only minor, non‑universal vulnerabilities that also exist in publicly available models. Key points from Anthropic’s response: - Strong safeguards were built into Fable 5, tested for thousands of hours with U.S., U.K., and private red‑team partners, and deemed more effective than prior models. - No universal jailbreak has been discovered; only narrow, low‑impact bypasses are known. - A defense‑in‑depth strategy, including 30‑day data retention for monitoring, is in place. - The cited jailbreak involved asking the model to analyze and fix a codebase—capabilities common to other frontier models. - Anthropic argues that recalling the models over a narrow issue would set a precedent that could halt future model deployments and calls for a transparent, fact‑based statutory process. Anthropic apologizes for the disruption and is working to restore access.

The piece argues that AI must remain open‑source to preserve operational freedom and prevent dependence on a limited set of closed institutions. It emphasizes that intelligence, as a civilizational infrastructure for work, education, science, creativity, public services, and national capacity, should be study‑able, build‑able, repairable, auditable, adaptable, teachable, preservable, and runnable without external permission. Open‑source AI is defined as usable, understandable, reproducible, locally deployable, economically viable, and community‑governed, resilient to changes or disappearance of dominant labs, hardware vendors, cloud platforms, or model providers. The text warns that concentration of control in a few frontier labs and platform companies could turn cognition into a subscription economy. It calls for American capacity aligned with global open standards to ensure freedom to run, inspect, modify, benchmark, teach, and preserve AI infrastructure.

Renault Group’s electric motor portfolio, branded as EESM, spans three generations without rare‑earth materials. - First‑generation motors (part 5A) debuted in the Kangoo Z.E (2011) and Zoe (2012) with 57‑100 kW output; the final 5AL version powered the 2020 Twingo Electric at 60 kW. - Second‑generation (part 6A) entered production in 2021; the 6AM motor equipped the Megane E‑Tech (early 2022) delivering up to 160 kW, later used in the 2024 Scenic E‑Tech and Alpine A290. - Third‑generation launch began October 2024 with the Renault 5 E‑Tech (6AK motor, 110 kW) and Renault 4 E‑Tech (March 2025). The Alpine A390 (Sept 2025) adds a front‑axle 6AM motor plus a twin‑motor rear set, totalising roughly 345 kW (≈470 hp). All motors are produced at Renault’s Cléon plant, with a next‑generation EESM line planned for 2027.

None

The page “putt.day” presents a daily mini‑golf feature. This entry, numbered 31, was published on June 12 2026 and introduces a single new hole for that day. The only visual element described is an image whose alt text reads “A floating mini golf course over water,” indicating that the depicted hole is set on a platform that appears to hover above a water surface, creating an aquatic‑themed mini‑golf experience. No additional text, rules, or commentary accompanies the entry. The content therefore consists of the series title, a brief description (“A daily hole of mini golf”), the specific episode identifier and date, and a single image caption describing the floating course.

Depthfirst’s autonomous security agent analyzed FFmpeg’s ~1.5 M‑line codebase and produced 21 zero‑day bugs for roughly $1 k, a tenth of the cost reported by Anthropic’s Mythos scans. Eight findings received CVE‑2026‑39210 through ‑39217 and involve heap, stack, integer and underflow errors across components such as the TS demuxer, swscale, VP9 decoder, DASH demuxer and option parser. The remaining 13 issues, tracked internally as DFVULN‑127 to DFVULN‑119, affect RTP depacketizers (AV1, AV1 AV1, AV1 LATM, MPEG‑4, JPEG, etc.), AVIF overlay, CAF, AVI, RTSP/RTMP parsers, and the option map logic; all are heap or stack overflows or integer over/underflows, many dormant for 10‑23 years. A highlighted exploit is a heap overflow in the AV1 RTP depacketizer (libavformat/rtpdec_av1.c). Skipping a Temporal‑Delimiter OBU advances the write cursor without allocating memory, allowing an attacker‑controlled 183‑byte RTP packet to overwrite the `AVBuffer.free` function pointer (offset 152) and trigger arbitrary code execution when the packet is freed. The vulnerability is reachable via a standard `ffmpeg -i rtsp://attacker/stream` command, requiring no special flags. All findings include reproducible proof‑of‑concept inputs confirming exploitability.

Apple rewrote its TrueType hinting interpreter in Swift for the Fall 2025 releases, replacing a legacy C implementation. The new interpreter provides full memory safety, maintains binary compatibility and pixel‑identical rendering, and runs on average 13 % faster. Development relied on extensive testing: a unit‑test suite covering 99.7 % of code and a minimized corpus of 4,200 PDFs (27 M glyphs) fuzzed to verify bitmap output against the reference interpreter. Performance gains were achieved by eliminating ARC and exclusivity overhead through copyable value types, using Swift 6.2’s `Span` for efficient sequence access, and avoiding data copies via projection types that safely wrap C structures. Hot‑path optimizations included in‑place stack operations with continuation‑passing to remove heap allocations. The top‑level interface remains an `@objc` class for Objective‑C++ interoperability. The source code is released on GitHub as a production reference, and the project’s lessons have been codified for LLM‑assisted C/C++‑to‑Swift migrations.

A local coding agent was built on macOS (Apple M1 Max, 64 GB) using llama.cpp compiled with Metal, the Gemma 4 26B‑A4B model in GGUF format, an Q8 MTP draft model for speculative decoding, and the Gemma 4 multimodal projector to enable image input. The baseline llama.cpp run delivered ~58 tokens/s; adding the MTP draft and tuning `--spec-draft-n-max` to 3 raised throughput to ~72 tokens/s (≈24 % gain) without slowing prompt processing. Benchmarks used a 128‑token Python‑diff task. llama.cpp outperformed MLX for this setup. The server is launched with `llama-server` specifying the main model, draft model, projector, speculative type `draft-mtp`, and context size 65536, exposing an OpenAI‑compatible endpoint at http://127.0.0.1:8080/v1. Pi (the terminal coding agent) is configured with a local provider pointing to this endpoint, supporting both text and image inputs. An alternative Qwen 3.6 35B‑A3B stack is provided; it offers better coding ability but slower performance (~55 tokens/s). The final stack includes the main GGUF model, MTP draft, projector, `--spec-draft-n-max 3`, and runs via tmux.

The article argues that a computer‑science (CS) degree remains a strong entry point despite recent headlines about AI‑driven job cuts. Federal Reserve data show 6.1 % unemployment for recent CS graduates (7.5 % for computer engineering), but when underemployment is included, engineers rank below a 20 % underemployment rate versus a 42 % average across all majors. “Entry‑level software engineer” postings rose ~47 % from late 2023 to late 2024 while actual hires fell ~73 %, creating “ghost jobs” that obscure real opportunities. Practical advice includes: - Leverage personal networks; ~26 % of offers come via referrals. - Target startups where junior hires share risk and gain broad exposure. - Create demonstrable experience (deployed projects, open‑source contributions) rather than relying on “toy” work. - Acquire hands‑on AI engineering skills (RAG pipelines, embeddings, vector databases, multi‑agent systems) beyond tool fluency; AI‑related roles grew 163 % in 2025. - Focus on durable systems‑thinking abilities, as demand for such engineers persists despite hiring fluctuations. The piece also notes upcoming Big‑Tech layoffs linked to AI, a robotics‑policy shift restricting Chinese components, and related industry news.

John Scott‑Railton reported that some malware authors have inserted references to nuclear and biological weapons into their spyware code. The purpose of these strings is to trigger safety refusals in large‑language‑model (LLM)–based security scanners, causing the tools to block analysis of the malicious payload. By prompting an LLM’s content‑filtering mechanisms, the attackers aim to evade detection and forensic examination. The tweet highlights this technique as a concrete example of how over‑reliance on first‑order safety checks can be exploited, emphasizing the need for more robust, context‑aware AI security solutions. The message includes a link to a source for further details.