How to secure and audit outbound traffic in an enterprise environment?

What follows is a long-winded background to preface the following question: What is the industry best-practice (or what recommendations would you give) for securing outbound traffic in an enterprise environment?

Background

We have a pretty typical enterprise environment: Linux and Windows hosts on unroutable IPv4 addresses behind a firewall/router/proxy. Among other things, these hosts run our application and database servers for our company's core service, which we develop in-house. The database servers have been locked down pretty extensively. Their addresses do not route (even with NAT/PAT) out of the internal network.

The application servers, and the servers that build and prepare the applications, on the other hand, require some connectivity to the outside world. These hosts need to gain access to Internet resources for a variety of reasons:

• integrate with public or private services,
• pull libraries from open-source repositories,
• download platform updates,
• get resources for ad-hoc troubleshooting,
• transmit statistics to affiliate monitoring systems,
• possibly other uses not yet identified.

Often these Internet resources can be identified by a host name or domain name, but it's not often they can be referenced simply by a destination IP address. The resource may be a specific subset of a resource, such as an application in a domain (graph.facebook.com) or a path on a host (google.com/a/company). We would like to be able to identify the resource as specifically as possible so as to avoid being overly permissive.

Our goal is to maintain a secure network. In particular, we want to:

• prevent or limit data exfiltration by a clever adversary who has gained unauthorized access to a system,
• monitor and account for activity originating from inside the network.

Our focus is on traffic originating from inside the network and terminating outside our secure environment. Furthermore, we aim to keep the permissions as tight as possible, specifying the source based on a class of host and the destination based on the resources that class of host requires.

Whatever we ultimately use, we would like to mechanize the process of granting access as part of the host provisioning process. Another requirement is that application development should be minimally impacted; the solution should look as much as possible like a simple TCP/IP network for authorized communication.

To this end, we have a few proposals on the table, some of which we've tried and some of which we may evaluate:

DMZ Firewall

The Firewall sits on the DMZ and routes traffic (network layer) based on a whitelist of allowable destination hosts. It defines a whitelist of IP addresses. The firewall simply discards traffic that doesn't correspond to the appropriate rule.

Advantages

• Applications can be written naturally without any consideration for the network.
• Supports transports other than HTTP/HTTPS.

Limitations

• Low granularity - An IP name or address could serve many applications and almost always serves multiple paths on that host.
• It's sometimes possible to resolve a hostname to a list of IP addresses using DNS and mechanically update that list, but the updates don't happen in real time.
• Rejected traffic is indistinguishable from a faulty network. Failures can be time-consuming, taking 15-60 seconds (or longer) to resolve as failed.
• Mechanization of the firewall is undesirable due to potential for abuse/failure.

HTTP Proxy Server

The proxy server, like the firewall, resides in the DMZ with connectivity to both the internal and external network. All outbound traffic must pass through the proxy server, which contains a white list of authorized resources by URL or partial URL. It only allows traffic that matches a specified resource on the whitelist. The proxy server operates at the HTTP/HTTPS protocol level.

Advantages

• Robust definition of allowed destinations.
• With a little host configuration, works naturally for many applications.
• Rejected traffic can usually be identified quickly.

Limitations (some apply specifically to our Stingray appliance)

• Applications must be aware of the proxy server and direct traffic to it.
• Some libraries require special handholding (or even bug fixes) to work properly in this environment. It's often difficult to tell in advance which libraries will be affected.
• Rules cannot differentiate easily based on source host class.
• Difficult to mechanize.
• Only works for HTTP/HTTPS.
• The network environment of the proxy can differ from the network environment of the host, leading to difficult-to-diagnose situations. For example, the application host can resolve hostname but proxy cannot, so the proxy returns a 404 response which is difficult to distinguish from a 404 response from the intended resource.
• The proxy cannot inspect encrypted traffic, and therefore cannot filter encrypted traffic any better than a firewall.

Advanced Perimiter Security Device

We are currently considering a device such as the Palo Alto Enterprise Perimiter. This device, like the others, would filter traffic. This device would do deep inspection of application-level transmission. It can inspect HTTP headers and manipulate traffic accordingly. It can even intercept and decrypt secure packets (SSL).

Advantages

• It's a commercially-supported, feature-rich approach.
• Deep packet inspection provides plenty of detail to apply fine-grained permissions.
• Applications conversing in plain text need not be aware of the device.

Limitations

• For us, it's a new investment and yet another appliance to learn/configure/manage.
• If SSL inspection is enabled, it violates the chain of trust. Applications properly configured for high security will balk or fail, so must account for the specialized environment.
• It's an unknown quantity. It's unclear if it will have the interfaces that would enable the mechanization we desire.
• A new capital expenditure.

The Question

What is the industry best-practice (or what recommendations would you give) for securing outbound traffic in an enterprise environment? Based on the details given above, is our stance too aggressive (or too lenient)? We're about to invest in one of these solutions (or maybe another) by developing tools to mechanize our processes, so any thoughtful advice on the best approach will be most appreciated.