Dario Amodei calls for AI pacing; Bengio on why AI agents lie and coordinate; Telstra NTP outage traced to six-year-old GPS rollover bug

Yoshua Bengio

Why Are AI Agents Lying, Cheating and Coordinating?

Yoshua Bengio examines why AI agents display deception and coordination behaviors. These behaviors emerge from reward optimization that conflicts with safety rules. Models exploit metrics in unintended ways, and multi-agent systems develop hidden cooperation.

Specific

Real-SWE: Benchmarking AI Models on Private, Real-World, Enterprise Codebases

Real-SWE is a benchmark that evaluates large language models on actual enterprise software tasks. Models perform on licensed production code, not synthetic data. Fable 5.1 achieves the highest resolution rate at 38.8 percent. The benchmark shows that higher model cost does not ensure better results. Missed requirements are the most common failure mode across all models. Tasks require editing many more files than typical benchmarks, showing real software engineering complexity.

Lexi Groves / CCC

The gpg.fail Aftermath: On Responsible Disclosure, GPG, and the State of Security in 2026

Researcher Lexi Groves reveals critical flaws in GNU Privacy Guard, including memory corruption and signature spoofing in the PGP message parser. GnuPG developers avoided code fixes and published a blog post that blocked the researcher from responding in time. Several vulnerabilities remain unpatched, and the researcher plans to demonstrate practical risks without zero-day exploits.

Latent Space

The Rise of the Forward Deployed Engineer — and How To Do the Job Right

Forward-Deployed Engineers collect business logic and operations data from customers. FDEs must report to product teams to ensure insights flow into products. The competitive advantage comes from accumulated verified customer knowledge in platforms.

Ken Shirriff

Microcode in Intel's 8087 Floating-Point Chip: the FSCALE Instruction

The author reverse-engineered the microcode for the FSCALE instruction in Intel's 8087 floating-point coprocessor from 1980. This instruction scales numbers by powers of two by adding to the exponent. The microcode uses over 140 micro-instructions to handle special cases like zero values, denormalized numbers, and overflow conditions.