Attackers now have powerful AI on their side. AI agents can scan, plan and run intrusions at machine speed, and they improve every month. The result is more attacks, more often and at global scale. In that environment, more companies will face a security incident at some point. What separates them is how they respond. Machine-speed threats call for a modern playbook with 4 pillars. Detect and contain automatically. Communicate fast and securely. Eradicate thoroughly. Keep shrinking the blast radius. This article shows the trend through 6 months of incidents. It then looks at a few companies, with Fireworks AI as a model of mature, structurally sound response.
TL;DR
- The trend: AI-powered attackers are driving more frequent attacks at global scale.
- The evidence: Several AI platforms disclosed security incidents between May and November 2026.
- Pillar 1: Detect and contain automatically. Human-speed shutdowns are too slow.
- Pillar 2: Communicate fast and securely, over out-of-band channels.
- Pillar 3: Eradicate thoroughly, with outside forensic experts.
- Pillar 4: Keep shrinking the blast radius. Data that is never stored cannot be stolen.
- Case studies: Fireworks AI and Klue show these pillars in practice.
Why are attacks becoming more frequent?
AI has changed the economics of attacking. Tasks that took skilled people days can now be delegated to agents.
In November, Google Threat Intelligence Group and Mandiant said AI misuse had reached a “critical inflection point”. They report that attackers use AI to shorten timelines and run multi-stage intrusions at machine speed. CrowdStrike reported that 1 group used AI agents against more than 200 organizations in 11 days. Palo Alto Networks’ Unit 42 found attacks moving about 4 times faster than a year earlier.
The shift started earlier. In 2025, Anthropic disrupted a campaign where AI did 80 to 90% of tactical work with little human input.
The tools keep improving, so the pace will keep rising. AI infrastructure is an especially attractive target, because AI access itself has value. The last 6 months show what that looks like.
What was disclosed in the last 6 months?
Incidents appeared across every layer of the AI stack. Together they show how broad the activity has become.
May: package registries. On May 10 and 11, the “Mini Shai-Hulud” campaign altered about 170 npm packages and some PyPI libraries. Affected projects included AI SDKs from Mistral AI and Guardrails AI. The NHS cyber alert lists the specific versions to avoid.
June: SaaS integrations. Klue, a market intelligence platform, disclosed misuse of OAuth tokens it held for customers. See case study 2.
August: self-hosted AI gateways. Microsoft described intrusions into exposed, unpatched LiteLLM, RAGFlow and Kestra deployments. Fixed versions exist for all 3 tools. CISA had listed the LiteLLM flaw as known exploited on June 8.
September: evaluation environments. Anthropic’s September threat report described attempts to obtain AI keys through software vendors. The Cloud Security Alliance summarized it. In 1 case, instructions were injected into a vendor’s automated evaluation sandbox.
October and early November: developer platforms. At least 9 AI developer and hosting platforms disclosed incidents involving stored customer secrets in October. Wiz’s tracker later listed further disclosures, for 16 platforms in total. A tampered LiteLLM release, version 1.107.0, was also identified, and users were told to roll back.
November: shared web vendors. On November 16, an attacker abused a flaw in Optimizely’s account signup flow to inject code into customer projects, reaching 178 organizations’ websites, according to Report URI, including HSBC, Salesforce, the BBC and Ledger. The code swapped crypto wallet addresses and skimmed payment forms; Optimizely removed the payload and fixed the flaw. Hudson Rock traced similar code through three other web services, beginning November 13. It is the same concentration pattern, one layer below the AI stack.
| Date (2026) | Platform | What was affected | Entry point, as reported |
| May 10 to 11 | npm and PyPI packages, incl. Mistral AI and Guardrails AI SDKs | Developer and CI credentials on installing hosts | Poisoned package releases |
| June 11 to 12 | Klue | Customer Salesforce and Gong data | Legacy credential, then stored OAuth tokens |
| August | LiteLLM, RAGFlow, Kestra deployments | Model provider keys, backend access | Exposed, unpatched gateways |
| October 6 | Fireworks AI | Environment variables in evaluation jobs, limited accounts | Misused internal credentials |
| October 14 to 15 | Zeabur | 3,218,965 environment variable records, 321,810 accounts | Long-lived admin tokens, exposed admin interface |
| October 15 | Lightning AI | Credentials; Studio taken offline | Credential access |
| October 16 to 17 | OpenCode | OpenAI, Anthropic and Google keys of about 1.4% of Console users | Legacy system |
| October 20 to 21 | Railway | Encrypted environment variables, some decrypted | Privilege escalation in build infrastructure |
| October 21 | LiteLLM 1.107.0 | Cloud credentials and LLM keys on installing hosts | Release pipeline compromise |
| October 23 | Warp | Agent decision logs and telemetry, about 1% of users | Third-party vendor dependency |
| October 23 to 24 | Fireworks AI | Billing data; for some customers, datasets and incomplete model weights | Incident at a third-party cloud provider |
| October 29 | Cursor | Altered extension copy in its marketplace cache, up to 5,000 users | Credential from exposed internal data |
| October 30 | Northflank | Customers told to assume environment variables were decrypted | Not disclosed |
| November 8 | NEAR AI | Inference server; 2 API keys exposed in logs | Remote code execution flaw in SGLang |
Selected incidents: Dates follow the linked sources, which sometimes differ by a day. Several October entries rely on compilations by Wiz and Helix rather than primary vendor disclosures.
What do these incidents have in common?
There was no single vulnerability. According to the Cloud Security Alliance, entry points varied. The common target was stored environment variables, often holding AI provider API keys.
AI keys carry real value, which explains the interest. CSA, citing Anthropic, notes that a key can be resold, used for compute or used as cover. Wiz estimated that misuse of 2 keys from these incidents exceeded $100,000 at list price. That is why spending caps matter.
Concentration is the other shared factor. A platform that stores secrets for many customers naturally draws attention. CSA describes this as the platform-chain pattern.
What does a modern incident response playbook look like?
No company can promise it will never be breached. Every company can control how it responds. Against attackers that move at machine speed, that response rests on 4 pillars.
1. Detect and contain automatically: A human-speed shutdown is no longer enough. An AI-driven intruder can move between systems in minutes. Mature teams wire detection directly to containment. That means kill-switches, automatic token revocation and host isolation. Wiz recommends alerting on bulk secret reads, the clearest signal in this year’s incidents. The best setups act on that signal automatically.
2. Communicate fast, and securely: Speed still matters most for revocation. Customers need to rotate keys, and upstream providers need to revoke access at the source. CSA notes that deleting a key from a platform does not revoke it. Mature teams also choose the channel with care. A Five Eyes joint advisory warns against communicating over the network under investigation. It advises out-of-band communication instead. CISA’s playbooks likewise call for out-of-band communications infrastructure as part of preparation. That keeps an attacker who is still watching from learning the response plan.
3. Eradicate thoroughly: Closing the hole that was used is only the first step. Top teams assume the attacker planted backdoors and hunt for them. They bring in independent forensic experts, such as Palo Alto Networks Unit 42 or Mandiant. The same advisory recommends third-party incident response support. It also warns against fixing symptoms instead of the root cause.
4. Keep improving by shrinking the blast radius: Attackers improve every month, so security is never finished. The strongest proactive defense limits what any future attacker can reach. Zero Trust access means no credential is trusted by default. Ephemeral architecture means sensitive data is not kept at all. If data isn’t stored, it can’t be stolen.
Some companies already apply these pillars. The 2 case studies below show how:
Case study 1: How did Fireworks AI respond?
Fireworks AI is an inference and training platform. It disclosed an incident in October, 2026, during a busy month for the industry. Its response is a strong example of mature, structurally sound incident response across the 4 pillars.
Pillar 1: detect and contain
Fireworks identified unauthorized use of internal credentials on October 6. A second event began on October 24, through a third-party cloud provider that hosted its R&D machines. Fireworks detected that activity and closed it on the same day it began. That is the kind of operational agility machine-speed threats demand.
Pillar 2: communicate fast
Fireworks began notifying affected customers on October 7, about 1 day after detection. It published a public bulletin on October 9 and has updated it as findings came in. It also notified law enforcement. The bulletin lists affected resource metadata and billing information. For certain customers, it also lists datasets, incomplete model weights and secrets.
Pillar 3: eradicate thoroughly
Fireworks engaged Palo Alto Networks Unit 42 for forensic analysis. Bringing in an independent, top-tier team is a mature decision. It means the investigation looks beyond the first entry point, for anything an attacker may have left behind. Fireworks also completed rotation of employee API keys by October 9.
Pillar 4: a smaller blast radius by design
Fireworks’ most important protection was already in place. It says it detected no impact to its Inference, Training SDK and Nexus products. All 3 run with zero data retention by default, so prompts and outputs are not stored at rest. That architecture kept the core products out of the reported impact. Fireworks has since gone further. It paused its Batch API, and its documentation says stored Batch inputs and outputs will be permanently deleted. Enterprise customers can also enforce an account-wide ZDR policy.
The philosophy behind it:
On November 13, 2026, Fireworks CEO Lin Qiao posted on X to explain the company’s data retention policy. She described 2 standard API designs:
- Customer-managed retention (upload, use, delete). A user uploads a dataset, runs jobs on it, then deletes it. Batch APIs are the typical example. The platform keeps the data until told to delete it, because it cannot know when the user is finished. Cleanup is the user’s job. Fireworks says it does not retain data after deletion.
- Ephemeral processing (send, compute, return). The platform processes a request and returns the result without keeping the content. This is zero data retention. Fireworks applies it to its inference APIs, its Training SDK, including serverless training, and Nexus. Prompts and generations are not persisted at rest.
She compared it to saving a spreadsheet in a cloud drive versus passing its numbers to a calculator. In both designs, customers control their data. Her conclusion is that ephemeral processing is preferable whenever the workflow allows it. It leaves no forgotten datasets or missed deletion calls, and every retained copy is something to protect. In our view, this is the gold standard for proactive defense, and the October events bear it out.