Anthropic · Cybersecurity · Critical infrastructure
The AI cybersecurity fire brigade has arrived. Somebody still has to repair the building.
Anthropic launched its Cyber Mission on October 8, combining resources for critical infrastructure and open-source software. Its Critical Infrastructure Defense Program brings Claude models, on-site engineers and threat research to eleven founding partners, including security providers, industrial manufacturers and consulting firms. The company says work with several partners is already underway. The practical challenge is securing operational technology without disrupting the machinery it controls.
OSS Scanner supplies a separate, opt-in route for eligible open-source projects: free periodic scans, explanations, reproducers and candidate patches where available. These outputs arrive without human review or triage. Anthropic expects a true-positive rate above 90%, but that is an expectation rather than a guarantee. Its research post reports an early expert review of 97 high- or critical-severity findings: 85 met its disclosure standard, eleven were real but duplicative, and one was invalid. That selected sample does not establish accuracy across every project and severity.
The useful measure is how many important weaknesses get safely repaired before exploitation. Maintainers still need to reproduce findings, assess severity, check patches and manage releases; operators of physical systems must also protect continuity. Faster discovery can help, but it can create a queue faster than a small team can clear it. The AI found 47 problems. The maintainer now has 47 problems. Better equipment matters. So does giving the people using it enough capacity to finish the job.
Source: Cyber Mission scope and infrastructure partners ↗Source: Unreviewed reports, eligibility and early validation ↗
OpenAI · Safety oversight · Research governance
Three researchers were fired. The argument is about who gets to inspect the laboratory.
OpenAI said Friday that it fired three researchers last week after an investigation found violations of sensitive-information policies, according to Reuters. AP identifies them as Tomek Korbak, Jasmine Wang and Mikita Balesni. The researchers warn that the dismissals could chill safety criticism and external oversight. OpenAI denies that the firings were about raising safety concerns. These are competing accounts, rather than an independently resolved explanation.
Korbak says his dismissal concerned how he communicated with METR, the independent evaluator involved in investigating the Hugging Face agent incident. OpenAI has not publicly supplied enough specifics to assess its allegation. Confidentiality obligations can be legitimate, and a claim of retaliation is not proof of it. Equally, meaningful independent evaluation requires reliable access to relevant evidence. The issue is whether those responsibilities can coexist under rules that everyone involved understands.
That is an institutional engineering problem with technical consequences. External reviewers need a dependable process for investigating incidents, and employees need to know how authorized disclosures work. A system that relies on informal understandings becomes fragile when an uncomfortable finding arrives. Friday establishes a serious dispute about that process, without settling who violated which rule. The laboratory can build increasingly capable intelligence. It still needs a trustworthy way to let somebody check the work.
Source: Reuters on the company’s statement and timing ↗Source: AP on the researchers’ account and OpenAI’s response ↗
Firmus · AI infrastructure · Investment
The AI infrastructure boom met an investor who asked to see the finished building.
Australia’s Nvidia-backed Firmus shelved its planned roughly US$5 billion IPO on Friday. The initial A$11 share price implied an equity valuation of about US$30.6 billion, nearly three times its August fundraising valuation. Firmus says it will seek private capital and consider alternative international public-market options. The company cited market volatility and conditions; Reuters reports lukewarm demand and investor concerns about debt, valuation and expansion risk.
Reuters describes two leased, operating data centers in Melbourne and Singapore, with five more planned across the Asia-Pacific. It also reports that orders weakened after CDC Data Centres’ chief executive said their planned 1.6-gigawatt development with Firmus was no longer underway. That is a more specific problem than investors suddenly deciding AI has no value: the proposed price depended heavily on construction and financing assumptions they were being asked to accept.
Functioning compute infrastructure can be valuable while a particular valuation remains difficult to justify. The withdrawal does not establish that Firmus has stopped operating, canceled every expansion plan or proved an industry-wide collapse. It shows that financing the future still requires someone willing to pay on the offered terms. The distinction feels overdue in an industry fond of counting tomorrow’s capacity today. The future can be enormously valuable. That does not mean every invoice addressed to it should be paid immediately.
Source: Reuters on the withdrawal, valuation and execution risks ↗
