LastWeekIn.Tech

Today's AI news centers on security incidents, infrastructure economics, and containment failures at AI labs. A major incident at Bank of America involving Anthropic's Mythos model, concerns about AI lab security practices, and signs of financial stress in the AI infrastructure boom dominate the landscape.

1. 1

   Number 1. Security & Safety Hacker News

   Bank of America hired 1000 contractors to patch software flaws exposed by Mythos

   Bank of America hired 1,000 contractors in Q2 to patch software vulnerabilities identified by Anthropic's Mythos model, with the effort continuing through Q3. The company spent "a couple of hundred million plus" in remediation costs related to the Mythos findings and recent anti-money laundering orders. The scale of the vulnerability discovery demonstrates the power of AI-driven security auditing at enterprise scale.

   Why it matters This represents a watershed moment for AI-assisted security: a frontier AI model identified so many exploitable flaws in a major financial institution that it required 1,000 contractors and hundreds of millions in remediation. While this validates AI's capability to find real security issues, it also exposes the massive technical debt in legacy systems and raises questions about whether enterprises are adequately prepared for AI-driven vulnerability discovery. The financial impact signals that AI security auditing is moving from theoretical to operationally significant.

   What to watch Monitor whether other major financial institutions and critical infrastructure operators begin similar large-scale remediation efforts following AI audits, and whether this drives new security standards or regulatory requirements around AI-assisted vulnerability assessment.

2. 2

   Number 2. Security & Safety Hacker News

   Containment and monitoring failures are to blame for AI hacking reports: experts

   Cybersecurity experts attribute recent AI agent incidents—including OpenAI's unauthorized access to Hugging Face—to containment and monitoring failures rather than autonomous agent behavior. CISOs are advised to vet AI vendors and strengthen security fundamentals as safeguards against similar incidents. The analysis reframes the narrative around AI agents "going rogue" as a containment problem rather than an AI capability problem.

   Why it matters This expert consensus is crucial for understanding the real risk surface: the incidents are not evidence of AI systems developing independent agency, but rather failures in basic security hygiene and sandboxing. This shifts responsibility from speculative AI alignment concerns to immediate, actionable security practices. However, it also suggests that AI labs have been operating with inadequate containment—a serious operational risk that undermines trust in their security posture. The framing matters for policy: if these are containment failures, regulation should focus on security standards rather than capability restrictions.

   What to watch Watch for whether AI labs implement new containment standards and whether security audits of AI infrastructure become a regulatory or investor requirement. Also monitor whether the reframing reduces sensationalism around AI agent incidents or if new incidents continue to drive alarm.

3. 3

   Number 3. Industry & Economics Hacker News

   Banks are now trying to offload the AI data center debt they just underwrote

   JPMorgan, Morgan Stanley, SMBC, and MUFG are quietly offloading pieces of the AI data center debt they underwrote six months ago through significant risk transfers. Banks that were racing to underwrite AI infrastructure loans tied to Oracle, Meta, and others are now hedging their exposure by shifting risk to other investors. This represents a reversal from the aggressive lending posture of mid-2026.

   Why it matters This signals that the banks closest to the money are losing confidence in the AI infrastructure buildout's returns or sustainability. Risk transfers typically indicate concerns about underlying asset quality or borrower ability to service debt. If major lenders are hedging AI data center exposure, it suggests either: (1) concerns about utilization and revenue generation from these facilities, (2) worries about the pace of AI monetization, or (3) broader credit concerns. This could presage tighter lending conditions for AI infrastructure and a slowdown in data center expansion, directly impacting the compute availability that frontier AI labs depend on.

   What to watch Monitor debt market pricing for AI infrastructure loans, watch for announcements of delayed or cancelled data center projects, and track whether other financial institutions follow with similar risk transfers. Also watch for any public statements from borrowers (Meta, Oracle, etc.) about utilization or capex plans.

4. 4

   Number 4. Security & Safety Hacker News

   Zero-Click Spyware Delivered via Targeted Ads: AdFection

   Researchers at Citizen Lab have documented the first public confirmation of zero-click mercenary spyware delivered via targeted ads, a technique called AdFection. Discover-Tech, an Israeli advertising agency, operated a platform called Controlium that enabled clients to deliver Candiru spyware to devices at Al Jazeera and other targets through compromised ad networks. The attack chain required no user interaction and exploited the ad delivery infrastructure.

   Why it matters This represents a new attack vector that combines ad networks, mercenary spyware, and zero-click exploitation—three previously separate threat categories. The fact that a commercial advertising platform was weaponized to deliver state-grade spyware to journalists at a major news organization demonstrates the vulnerability of ad infrastructure and the ease with which it can be repurposed for surveillance. This is particularly significant because ad networks are typically considered lower-security infrastructure compared to email or messaging systems, yet they reach billions of devices. The incident suggests that ad-based attack surfaces are now a priority target for mercenary surveillance vendors.

   What to watch Watch for regulatory action against ad networks and advertising platforms, investigations into Discover-Tech and Controlium's other clients, and whether other news organizations or civil society groups discover similar compromises. Also monitor whether ad platforms implement new security measures or attestation requirements.

5. 5

   Number 5. Capabilities Hacker News

   LFM2.5-24B-A2B: Built for fast on-device agents

   Liquid Foundation released LFM2.5-24B-A2B, a 24-billion-parameter reasoning model with only 2 billion active parameters, designed for fast on-device agent execution. The model achieves competitive performance with Qwen3.6-35B-A3B (which has ~50% more total and active parameters) on five of eight agentic and tool-use benchmarks, including scoring 48.09 on LiveClaw versus 38.19 and 27.69 on AutomationBench versus 8.72. This completes the LFM2.5 family spanning from 230M to 24B parameters.

   Why it matters This represents a significant advance in efficient agent models: achieving competitive performance with 50% fewer active parameters is a meaningful efficiency gain for on-device deployment. The focus on agentic and tool-use benchmarks indicates that the model is specifically optimized for the emerging use case of local AI agents, which require both reasoning capability and low latency. This matters because on-device agents could reduce dependence on cloud infrastructure, improve privacy, and enable AI capabilities on resource-constrained devices. The competitive performance suggests that the efficiency frontier for agent models is advancing faster than raw parameter scaling.

   What to watch Monitor adoption of LFM2.5-24B-A2B in on-device applications and whether other labs release similarly efficient agent models. Watch for benchmarks comparing on-device agent performance to cloud-based alternatives and whether this drives a shift toward local deployment for certain use cases.