Anthropic assesses GLM-5.3's cyber capabilities and lax safeguards
Anthropic reports that Zhipu AI's GLM-5.3 can autonomously build end-to-end cyber exploits, matching the capabilities of Claude Mythos Preview. Unlike other frontier models, GLM-5.3 was released without meaningful safeguards to limit misuse. In simulated tests, attackers bypassed these safeguards 64% to 100% of the time using simple techniques. In contrast, similar attacks did not succeed against safeguarded Claude models. NIST's CAISI also found GLM-5.3 to be the most cyber-capable open-weight model released to date, lagging the US frontier by about four months. Anthropic notes that while these capabilities benefit defenders, the lack of safeguards significantly increases risks for malicious actors.
anthropic.com
·
HN discussion
Livenerf begins tracking Claude Opus 5.5 for post-launch performance drift
A new benchmark called livenerf has started collecting daily data to detect whether Claude Opus 5.5 becomes less capable after its release. The project addresses reports that Anthropic may quietly alter model behavior, such as reducing effort or swapping in a smaller model, by establishing a clean day-zero baseline. It uses a deterministic harness with frozen prompts and statistical analysis to measure drift over time. As of September 29, 2026, the project has completed six of its first ten baseline days. The first official assessment of potential performance changes is scheduled for late October, after a full 30-day collection period concludes.
github.com
·
HN discussion
McDonald's uses AI to set local menu prices
McDonald's is deploying artificial intelligence to determine menu prices at individual franchise locations. The system analyzes transaction data from millions of daily sales to estimate customer willingness to pay and generate specific price recommendations. This approach has widened price differences for identical items between nearby restaurants. While the company describes the tool as a voluntary aid for franchisees, internal documents indicate that compliance is increasingly expected under new business standards. The strategy aims to optimize corporate profits but has created friction with franchisees facing rising operational costs and potential antitrust scrutiny.
cnbc.com
·
HN discussion
Memory makers shift to long-term contracts, ending price cycles
GamersNexus argues that memory manufacturers have broken the industry's traditional pricing cycles by committing to long-term agreements with fewer, larger customers. The outlet notes that average prices for 2 TB NVMe SSDs have risen 137% and 32GB DDR5 kits 363% since last September. Micron, Samsung, and SK Hynix are diverting capacity away from consumers, while Amazon cites rising on-premises hardware costs to drive cloud adoption. The article concludes this structural shift will likely eliminate future price lows, leaving consumers with permanently higher floor pricing.
gamersnexus.net
·
HN discussion
Cloudflare applies to become a public certificate authority
Cloudflare has applied to join the root programs of Chrome, Apple, Microsoft, and Mozilla, marking its first step toward becoming a public certificate authority. The company also signed a definitive agreement to acquire an established root from GlobalSign to ensure immediate device compatibility. While Cloudflare is not issuing certificates yet, it plans to be among the first to support post-quantum certificates and targets its first Merkle Tree Certificates for early 2027. This move follows twelve years of providing free Universal SSL, aiming to diversify the global certificate supply chain.
blog.cloudflare.com
·
HN discussion
Deser proposes a heap-based alternative to Rust's Serde
A developer has revived Deser, a Rust serialization library designed to address architectural limitations in Serde. The project inverts Serde's type-driven recursion by having the format push events into a sink, keeping state on the heap to prevent stack growth. While Deser cannot support non-self-describing formats like protobuf, the author argues it offers a viable path to resolve Serde's stability constraints and corner cases.
lucumr.pocoo.org
·
HN discussion