Tech • AI • Robotics
Anthropic scrutiny, GLM-5.3 cyber test, Tesla Roadster delay
Briefing
Anthropic faces fresh governance scrutiny
New reporting put Anthropic under a harsher spotlight as the company grows into a systemically important AI lab. Attention centered on Cammy Clark, described as a close adviser to chief executive Dario Amodei despite holding no formal role at the company. The reports said she helped connect Eric Schmidt to Anthropic in its early days and had explored an AI-focused fund concept that did not proceed. The episode sharpens broader questions about informal influence, access and governance at top-tier model developers.
GLM-5.3 edges Mythos 5 on bugs
Z AI said its open-weight GLM-5.3 scored 84.5% on CyberGym, narrowly above a reported 83.8% for Anthropic's Mythos 5. The benchmark focuses on reading code, identifying vulnerabilities and confirming they are genuine. If independently validated, the result would mark a notable parity claim in a highly sensitive cyber capability. The figures, however, remain vendor-reported rather than externally confirmed.
Anthropic keeps exploit-generation lead
The headline cyber gap reappears on ExploitBench, where models must convert discovered bugs into working attacks. GLM-5.3 posted 54.4%, far behind Mythos 5 at 78.0%. That spread suggests Anthropic still leads materially in the more operationally dangerous stage of offensive security. It also implies that bug-finding parity does not yet translate into exploit parity.
Mythos 5 wins on attack speed
Throughput figures reinforced Anthropic's edge beyond single-score benchmarks. In time-limited attack-development tests, GLM-5.3 completed 105 tasks in two hours and 130 in six, versus 181 and 247 for Mythos 5. The margin indicates faster iteration as well as stronger exploit execution. For defenders and policymakers, that makes the capability gap look structural rather than cosmetic.
Z AI plans restricted open release
Despite the sensitivity of cyber-capable systems, Z AI said GLM-5.3 would be released publicly in roughly two weeks. The company said rollout would follow security assessments and stronger safeguards. That timeline keeps pressure on rivals as open-weight models push further into offensive-security territory. It also revives the unresolved policy fight over how much dangerous capability should be openly distributed.
Lema pitches agent-risk monitoring
Startup Lema emerged with a pitch aimed squarely at enterprise anxiety over autonomous software. The company argues that conventional monitoring catches only anticipated failures, while its system watches wider organizational context for anomalies. Examples include an agent sending tens of thousands of emails or triggering abnormal business activity. The message reflects a real market shift as longer-running AI agents create larger and costlier failure modes.
Orbital AI data centers hit reality
The idea of space-based AI computing promises near-constant solar energy and naturally cold operating conditions, but the economics remain punishing. Launch costs, hardware replacement limits, latency and scarce bandwidth all work against orbital data centers. Low-latency service would likely require thousands of satellites in low Earth orbit, compounding congestion and debris risks already highlighted by constellations such as Starlink. In practice, underused terrestrial power and compute infrastructure still looks far easier to expand.
Tesla Roadster delays stay in focus
Fresh discussion around Tesla's Roadster underscored how long the halo vehicle has remained in limbo. The renewed attention lands amid a wider week of scrutiny over execution promises, valuation narratives and product timelines across technology. While new specifics remain limited, the Roadster has become a recurring symbol of stretched deadlines at Tesla. Its continued absence matters because flagship products still shape investor confidence far beyond unit sales.
Videos covered
Tesla’s Roadster, Car Week Takes, Unwell Beverages Shuts Down | Diet TBPN
- •Startup pitches lean into cinematic anxiety
- •Anthropic leadership faces wider public scrutiny
- •Questions of privacy collide with AI power
China Finally Beats MYTHOS 5 With New GLM 5.3 (Plus Anthropic's New Model 2)
- •Benchmark claim
- •Exploit gap remains large
- •Speed favors Anthropic
"Musk vous vend du rêve pour lever des milliards" : le piège des data centers spatiaux — David Gurlé
- •A seductive idea with harsh physics
- •Scale would require thousands of satellites
- •Maintenance is a core economic weakness